Gemini Updates & Release Notes

New antibiotics are urgently needed to help with rising levels of antimicrobial resistance, but searching for new therapeutics is often a slow, labor-intensive process. With Google Antigravity, our agentic development platform, this process can now be made easier.

Google Antigravity provides an agentic science workbench that empowers biologists to discover, structurally evaluate, and prioritize promising targets quickly, compressing what used to be a fragmented search over multiple databases into a single session. It also helps researchers scan through millions of previously uncharacterized proteins in the UniProt database using ProtNLM, a model that predicts biological function from amino acid sequences.

This technical walkthrough shows how biologists used Google Antigravity to uncover a novel antimicrobial peptide candidate which has predicted activity against the pathogens involved in respiratory tract infections. The lead identified here is now undergoing wet-lab validation.

Finding new antimicrobial peptides

Rising antimicrobial resistance is one of the most pressing global health crises of our time and could result in an estimated 10 million annual deaths globally by 2050. To prevent this, we urgently need to discover novel therapeutics. Antimicrobial peptides (AMPs), particularly bacteriocins, offer a promising alternative to traditional antibiotics because they have naturally evolved to target fundamental vulnerabilities in competing bacteria, from cell envelopes to essential intracellular machinery, making it difficult for pathogens to develop resistance.

The exponential growth of genomic and metagenomic sequencing has populated global databases with billions of natural amino acid sequences. This pool of natural data has already produced globally used bacteriocins like Nisin, alongside recent genome-mined breakthroughs such as Darobactin, Dynobactin, and Lariocidin.

Many, however, remain locked within the unannotated "dark matter" of protein databases, with their function uncharacterized because they aren’t similar to known proteins. For example, the world's leading public resource of protein sequence and functional information, UniProt, holds millions of sequences with unknown functionality, transferred directly from genome sequencing projects.

To bridge this gap, Google developed ProtNLM (Protein Natural Language Model), a deep learning model trained to directly read the underlying evolutionary “grammar” of amino acid sequences. This allows the model to predict biological function directly from raw protein sequences, enriching millions of previously unnamed UniProt entries with high-confidence functional labels.

Overview of the ProtNLM pipeline translating raw amino acid sequences into natural-language biological annotations for unreviewed UniProtKB/TrEMBL entries. The functional annotation scope of the system covers protein names, function comments, subcellular locations, and GO terms, in addition to keywords.

Collaborative intelligence

Scientific discovery is rarely a single step process; finding viable candidate proteins requires human intuition and iteration to incorporate new empirical evidence. Working in Antigravity, the researcher and AI agent begin by setting a focused biological goal: find and computationally validate novel AMPs in Streptococcus pseudopneumoniae. During an interactive, multi-turn session the researcher steers the scientific strategy, deciding which questions to ask and what evidence to collect, while the agent handles the heavy lifting of cross-database orchestration, structural analysis, and literature synthesis.

Navigating a vast search space: The agent queries the database and filters millions of entries down to a handful of candidates worthy of attention. Here, the agent finds a bacteriocin (Uniprot ID: A0A0U0D591) with a length of 97 amino acids from the Lactococcin 972 family.

The candidate we uncover in this walkthrough represents a biologically plausible lead backed by structural and genomic evidence but it still requires wet lab validation. What makes this meaningful is how dramatically it lowers the barrier to everyday biological workflows, turning days of fragmented database triage into a single collaborative session. While this session focused on a narrow amino acid window in Streptococcus pseudopneumoniae, researchers can take the conversation in any direction, such as layering in new filtering criteria, exploring different sequence lengths, or targeting other priority pathogens. By broadening the search from human commensals to environmental microbes and beyond, scientists can guide the agent across UniProt to discover novel AMP candidates at scale.

A shared 3D canvas for interactive collaboration

A protein's 3D structure can help us understand its biological function. Examining a candidate’s predicted fold in space is key. To provide this end-to-end workflow in a single workspace, we are introducing the Biomolecular Structure Viewer, a 3D visualization tool embedded directly into Antigravity.

The researcher and agent work together over a shared visual state. As the agent retrieves structures from AlphaFold Database or the scientist’s own predictions, it can dynamically drive the viewer. For instance, it can highlight key functional regions to visually explain its hypotheses. Crucially, this awareness flows both ways: actions like highlighting specific residues or pinning notes directly onto a binding pocket are synchronised back to the agent. A researcher can simply mark a region of interest in 3D space and ask, for example, "How might this β-strand affect Lipid II binding?", allowing the agent to immediately reason over the exact structural context the scientist is looking at.

The Biomolecular Structure Viewer inside Antigravity showing the AlphaFold-predicted structure of the candidate bacteriocin, giving the researcher and agent a shared 3D canvas to inspect and annotate functional regions.

Accelerating Scientific Discovery

Historically, uncovering insights like those in the workflow demonstrated above required fragmented and specialised software: writing custom scripts to query databases, manually exporting files to 3D modelling programs, managing disconnected genomic tools and working with multiple experts to design the wet lab assays. Because wet-lab validation remains resource-intensive, narrowing the search space to only the highest-confidence leads upfront is critical. Using Google Antigravity allows researchers to navigate the vast biological search space efficiently while leveraging accurate protein function predictions from a domain-specific model.

Google Antigravity executes this end-to-end scientific reasoning in one place. This allows scientists to spend more time evaluating high-quality hypotheses, running experiments in the lab and solving research challenges.

Acknowledgements

We would like to acknowledge the contributions of the following team members: Maria Gabriela Cardoso, Victoria Johnston, Lexi Walker, Gena Hong, James King, Richard Green, Filip Radlinski, Artem Sobolev, Michael V. LeVine, Ruijie K. Zhang, Zachary Wu, Luca Schulz, Amy Danson, Tristan Oliver Kwan, Ryan Krueger, Oleg Kovalevskiy, Doug Fritz, Yunhan Xu, Sebastian Nowozin

Original source

Universal Cart launches in the U.S. today. Here are four ways it can help with holiday shopping — including organizing products from multiple stores into a single cart, tracking price drops, and more.

As we build for the future of commerce, our focus spans two fronts: unlocking agentic capabilities that act on your behalf, and designing practical tools that give shoppers confidence and control. Google’s Universal Cart is built for that second track — mirroring how people actually shop today across Google, without losing track of their items, their savings, or the merchants they trust.

The Universal Cart, which we previewed earlier this year, is an intelligent shopping cart and your new hub for shopping on Google. The cart puts you in control, helping to organize your shopping, keep track of items you love from across retailers, stay on top of promos, and enjoy effortless checkout powered by the Universal Commerce Protocol (UCP). Today, Universal Cart is rolling out in the U.S. across Google Search and in the Gemini app.

The cart for all your shopping

Here are four things the new Universal Cart lets you do to stay organized and shop more confidently this holiday season.

• Save products from multiple retailers, wherever you’re shopping on Google.

  Universal Cart offers a more organized alternative to tabs and screenshots. The cart saves your progress across Google so you never have to start your shopping from scratch. Let’s say you save a handbag from Steve Madden to your cart in AI Mode in Search, and later a sweater from Quince and lipgloss from Ulta Beauty in the Gemini app. Those products will be there in your cart when you come back later so you can pick up right where you left off.

• Keep tabs on pricing or inventory changes for your entire list.

  The cart is built on the 60+ billion continually refreshed product listings in Google’s Shopping Graph. This allows it to proactively work for you in the background, sending notifications about price drops or when an item comes back in stock.

• Identify hidden savings or points from your cart.

  The cart’s intelligent reasoning can help anticipate your needs. Say you have a couple different credit cards, but don’t know each one’s specific benefits. The Universal Cart is built on Google Wallet so it understands your card perks and can intelligently help you choose between payment methods. It’ll also show you relevant promo codes right from the cart.

• Make seamless, trusted purchases when you’re ready to buy.

  In online shopping, trust is everything. You can securely check out with Google Pay in just a few taps with many of your favorite stores right in the cart, or you can transfer your items directly to the merchant’s cart to finish buying on their site. Try these simplified checkout experiences with merchants like Nike, Sephora, Target, Ulta Beauty, Walmart, Wayfair, and many Shopify merchants. Either way, you will receive direct confirmations and customer support from the merchant, and they will remain the seller of record.

As the holiday season kicks into high gear, the Universal Cart is just one way we’re giving shoppers the convenience and smarts they need to make the best shopping choices.

Original source

Introducing Gemini 4 Argon: our frontier model for complex tasks

• What: Starting today, Gemini 4 Argon is rolling out globally for Google AI Pro & Ultra subscribers in the Gemini app. 4 Argon is our newest frontier model, built to take your most complex tasks from start to finish. Whether you are analyzing a complex financial report, creating a deck from scattered sources, or building an interactive 3D model, Argon brings frontier intelligence to your everyday workflow, turning your ideas into outputs you can use.
• Why: We want to make Gemini the most helpful AI assistant for everyone. With the frontier performance of Gemini 4 Argon, you can now turn your most ambitious ideas into real projects.

Code, prototype, and build with Google Antigravity in the Gemini macOS app

• What: Google Antigravity is now built into the Gemini macOS app, so go straight from brainstorming in chat to coding apps, building web experiences, and prototyping ideas. Spin up a coding agent in Antigravity, approve its plan, and head right back to your Gemini chat. The agent builds and tests the code in the background, notifying you the second it’s ready to review.

  To get started, click Antigravity at the top of the Gemini app on your Mac. Available in English for personal Google Accounts and eligible work accounts (18+) across all regions where Antigravity is supported.

  Download the Gemini macOS app at gemini.google/desktop.

• Why: We want to help you turn ambitious ideas into working code and prototypes faster, right where you already work on your Mac. With Antigravity built into the Gemini app, you get an agentic coding partner that plans, builds, and verifies complex code, making it easier to bring your projects to life.

An easier way to customize how Gemini works for you and more connected apps

• What: Open the left menu and click Customize to tailor how Gemini responds and connects to your everyday tools in one place. You can discover apps and add skills so Gemini can tackle your most repetitive tasks, like writing in your style or following your formatting rules.

  You can connect more than 40 apps or services to get things done in one conversation. Whether you are ordering groceries on Instacart, reserving a table on OpenTable, or editing photos with Adobe, you decide which apps to connect to and can manage your settings at any time.

• Why: We want to make Gemini work the way you do. By bringing your favorite apps and custom skills together in a single conversation, Gemini helps you skip the context switching and get more done, faster and easier.

  Customize is rolling out globally across web, mobile, and the Gemini desktop app on Windows and macOS, for all user tiers. Most connected apps are available in English in the U.S., with select apps available in additional English-speaking regions, Brazil, and India.

Original source

As previously announced, skills in the Gemini app are starting to roll out today, with an expected completion by mid-November. In the coming weeks, we’ll also introduce critical features like the ability to share Gemini app skills and an integration with Google Drive. We’ll post on the Workspace Updates blog as soon as those capabilities start rolling out.

We also want to share an update on the rollout of skills in Workspace. We previously shared that the launch to Rapid Release would be complete by October 12, 2026, and the rollout to Scheduled Release would begin on October 19, 2026. While we work to ensure skills in Workspace launch in a safe and reliable manner, those rollouts are slightly delayed. We will share the updated timeline here as soon as it’s confirmed.

Getting started

• Admins: Visit the Help Center to learn more about admin controls for skills in Workspace. There are no admin controls for skills in the Gemini app.
• End users: Visit the Help Center to learn more about skills in Workspace and skills in the Gemini app. Google Classroom users can also explore our guidance on creating and using teacher-led AI experiences directly in Classroom.

Rollout pace

• Skills in the Gemini app
  • Rapid Release and Scheduled Release domains: Gradual rollout starting on October 13, 2026, with expected completion by mid-November, 2026
• Skills in Workspace
  • Rapid and Scheduled Release domains: Delayed until further notice

Availability

• Skills in the Gemini app
  • Available to all Google Workspace customers and users with personal Google accounts
• Skills in Workspace
  • Business: Business Starter, Standard, and Plus
  • Enterprise: Enterprise Starter, Standard, and Plus
  • Education Add-ons: Google AI Pro for Education
  • Other Add-ons: AI Expanded Access

Note that skills in Workspace are available only to users over age 18; skills in the Gemini app will be available to users of all ages.

Resources

• Workspace Blog: Teach Gemini your team’s know-hows with skills in Google Workspace
• News from Google Blog: Let skills in Gemini tackle your most repetitive tasks
• Workspace Handbook: AI skills 101: Customizing AI from Google for your role
• Workspace Admin Help: Allow people to use skills in Studio and Gemini in Workspace
• Workspace Studio Help: Learn how skills work in Google Workspace Studio
• Gemini Apps Help: Create & manage skills for Gemini Apps
• Workspace Admin Help: About the transition from Gems to skills
• Education Transition Guide: Gemini Gems and Skills Transition Guide for Education

Original source

Learn more about our approach to safety and security for Gemini 4 Argon.

Anca Dragan
VP, AI Safety & Behavior

Raluca Ada Popa
Head of Gemini Security, Google DeepMind

Gemini 4 Argon is a step forward in what models can do — especially when it comes to deep work in complex, real-world tasks across domains including coding, finance, legal, cybersecurity, and science. Its cyberdefense capabilities are frontier-level and designed to give defenders an advantage over attackers. To maximize its potential, we first wanted cyber defenders to put it to work via our Fairwind Program, while our own security teams used it to scan critical systems inside and outside of Google.

This window of limited access showed why these capabilities matter. For instance, Wiz used Argon to uncover a critical bug in hospital software that previous models had missed, helping protect hundreds of thousands of patient records. Meanwhile, Google’s own security teams used Argon to find and patch bugs across internal software or open-source software. This included fixing foundational vulnerabilities affecting systems globally, which are going through our responsible disclosure process.

Today, we are beginning to roll out Argon broadly, as well as expanding our Fairwind Program to additional trusted cyber defenders. To do this well, we grappled with two big questions:

1. How do we build the right safety guardrails? If Argon can help defenders find complex software bugs or help scientists design experiments, how do we stop bad actors from using it to launch cyberattacks or create biological threats? (Misuse)
2. Can we trust the model to do the job the right way? As models become more resourceful problem-solvers, how do we make sure they don’t cut corners, cross boundaries, or make incorrect high-stakes decisions? (Misalignment and Prompt Injection)

Our full Safety and Security Report systematically covers our set of defenses. Below are some highlights from how we approached these two questions for Gemini 4 Argon.

Misuse prevention across high-risk domains

The same coding and reasoning improvements that enabled Argon to find and patch vulnerabilities also drove better performance on our cybersecurity tests — solving all 12 of the end-to-end simulated cyberattack scenarios, up from five in our previous generation. In biology evaluations, unguarded versions of the model also showed a much stronger ability to help troubleshoot complex lab procedures step by step. Following our Frontier Safety Framework, since we saw elevated capabilities in the cyber and biological domains, we put our next-generation set of frontier safeguards in place across both areas.

First, when it comes to potentially dangerous requests — from writing exploit code to helping with dangerous biological materials, Argon is significantly better at not fulfilling harmful requests than our previous models. On an adversarial set meant to stress test the model, violations reduced nearly twofold — also significantly better than previous Gemini models. Second, we wrap the model in three layers of real-time protection; this is to prevent determined attackers that might try to jailbreak (i.e. trick) the model by disguising a dangerous request as a harmless story or breaking it into dozens of small questions over several weeks. Here is how these protections work:

• Input screening: Smaller models check incoming requests and conversation context for malicious intent or dangerous requests.
• Internal introspection: Even if a request is cleverly disguised to look harmless, small transformers ("activation probes") read the model’s internal neural signals and are trained to spot when it is activating knowledge about cyberattacks or biological threats.
• Continuous monitoring: As the model reasons and writes its response across multiple steps, we combine these signals and take action when necessary such as cutting off the agent mid-stream if the responses become problematic.

Combining real-time protection, faster fixes for new jailbreaks and enhanced monitoring for recurring misuse has made our systems better at preventing real-world harm. For instance, external bio jailbreak red teaming could only unlock less than 1% of bio sub-tasks we test for.

Real-time misalignment monitoring

Misalignment can sometimes mean an agent deciding on its own to pursue a goal different to what was intended by the user. Because Argon is much better than earlier models at understanding the environment it is running in, we test carefully for that kind of behavior. For example, we ran realistic "honeypot" tests to see if the model would increase its ability to take covert actions, such as sabotaging safety monitoring. We didn't find any instances of that behavior.

But what we do see in practice today is something much more familiar: an agent trying too hard to finish the job it was given. When a resourceful assistant is told to get a task done and runs into a locked door, it can treat that obstacle as just another puzzle to solve.

This is why we trained Gemini 4 Argon to respect system boundaries, report negative findings honestly, and pause to ask for human confirmation. On coding benchmarks where models have no incentive to hold back, attempts to take a shortcut or cheat on the task dropped to just 2% on Gemini 4 Argon, our best performance for recent Gemini models.

And while this is a major step forward, when a model runs millions of complex tasks, even a rare misstep will eventually show up. As a second line of defense, we built an automated misalignment monitor that reads both the agent’s actions and its internal "scratchpad" (its step-by-step chain of thought) in real time, blocking severe boundary violations before they happen and flagging serious internal cases for human responders. (We continually harden our training and evaluation environments against potential escapes. You can read more about this in our agent control roadmap). We run this monitor in our products, on our internal coding agents, during our training runs, and during our capability evaluations.

When training tasks were difficult or when training environments were flawed, our monitor caught the model, in rare instances, creatively attempting to game the test. The model attempted to escalate container privileges to alter grading scripts, read answers from the test runner's memory, and in one case it even crashed the evaluation sandbox to avoid getting scored for a failed run. While we use these catches to fix flawed tasks, we do not use them to otherwise intervene with the training process to avoid overfitting to the monitor.

And we always follow one rule: we don’t penalize the model for what it writes in its chain of thought during training. Here’s an example to illustrate why: in a reinforcement learning training run where a shortcut penalty was accidentally applied to the chain of thought, openly discussed shortcuts dropped by 61%, but that did not mean the model took less shortcuts. Instead, the share of shortcuts the model took silently — without mentioning them in its thoughts — more than tripled (from 11.2% to 36.4%). Avoiding thought penalties encourages the model’s reasoning to stay transparent and monitorable.

Prompt injection resistance

To trust an agent to do the job right, we need to ensure that malicious content the agent comes across cannot trick it to take a harmful action. When agents browse webpages or read code, they can encounter hidden, malicious instructions, or context designed to hijack their task. Through built-in adversarial training and real-time security checks, Argon leads in prompt injection robustness on Gray Swan’s Indirect Prompt Injection (IPI) benchmark, cutting attack success across 15 adaptive attempts to 0.7%.

The above safety and security defenses represent only three categories of our extensive defense-in-depth stack. To learn more about our approach to safety testing and mitigations, read our full Gemini 4 Argon Safety and Security Report and see the Gemini 4 Argon Model Card.

Original source

Gemini 4 Argon marks a step change in physical understanding, code generation, and visual-semantic reasoning—unlocking new workflows across the robotics stack and accelerating progress toward physical AGI.

You can see this leap in capability on the Embodied Reasoning (ERQA) benchmark, which tests a model's ability to understand and solve physical, spatial problems in 3D environments. Using pure reasoning, Argon scores 87%, outperforming every other frontier model.

As a general-purpose physical reasoning model, Argon can be applied across the entire robotics lifecycle. In this post, we share representative prompts for Argon across four key areas:

• Robot control → Mobile manipulation with MuJoCo practice, zero-shot robot control, and closed-loop steering of video world models.
• Real2Sim → Reconstructing physically calibrated simulation twins from robot camera feeds, egocentric human video, or hand-drawn paper sketches.
• Physical understanding → Agentic multimodal hindsight labeling including detailed understanding of dexterity, dense value functions and progress estimation.
• Research and self-improvement → Action improvement via simulation and iterative self-improvement on hardware.

Unlocking robot control with Argon

1. Spot mobile manipulation: Putting in the trash bin with MuJoCo practice

Argon controls a Boston Dynamics Spot® quadruped (12-DoF legs + 6-DoF arm) to learn a new skill: stepping on a foot-pedal trash bin to throw away soda cans. Argon creates an articulated MuJoCo digital twin of a foot-pedal trash bin to practice its moves before stepping on the physical pedal and dropping soda cans into the open bin.

Prompt:

There's a Diet Coke can behind you and another on the black table on your left.
Use the spot-mobile-manipulation skill to pick them up and trash them in the black foot-pedal bin.
Before stepping down on the pedal, verify in the front cameras that your front-right foot target is centered on the pedal, and use a 2-agent verification swarm (foot-verifier and hand-verifier) to verify both the foot-pedal step and the 3D hand-in-bin alignment before releasing.

2. Zero-shot control

Argon controls a simple robotic arm to place a fallen mug upright and put the crumpled napkin inside. Argon reasons with a visual coordinate harness and action primitives to perform this task over several turns.

Prompt:

[Attached: wrist_cam.png, obs.json]

Task: Get the robot arm to reorient the mug upright and put the napkin inside the mug.

Execution Discipline

• Simply output 1 sentence observation followed immediately by your curl command (with # Plan: on Turn 1 and # Thought: on every turn). Everything needed is in obs.json and the calibrated reticle/grid drawn on wrist_cam.png.

Visual Coordinate Reference (Annotated on wrist_cam.png)

• +X (UP in image): Away from robot base | -X (DOWN): Towards robot base
• +Y (LEFT in image): World left | -Y (RIGHT): World right
• +Z: Height above table | Yellow numbers: World (X, Y) in centimetres (2 cm grid at grid_z)
• Cyan + crosshair: Current 3D pinch center
• Cyan [ + ] brackets: Current jaw opening and closing line
• Green line: Current object long-axis orientation (yaw_deg: 0° = vertical along +X, 90° = horizontal along +Y)
• pitch_deg = 90°: Straight down (-Z) top-down grasp/place; pitch_deg < 60°: Tilted coaxial approach along finger axis

Closed-Loop Turn Workflow

1. Inspect current state & view wrist_cam.png:
   curl -s http://127.0.0.1:8081/obs (returns ee_pos, holding, status, workspace, wrist_cam_path).
2. Select VLM call granularity upfront on Turn 1:
   • Coarse (2–4 calls): look_at -> grasp -> place -> done
   • Pre-Grasp Hover Verify (4–7 calls): look_at -> goto(z=0.04) (verify cyan [ + ] brackets straddle target) -> grasp -> place_until_contact -> done
   • Fine Visual Servoing (8–20 calls): look_at -> grasp -> pre-pose goto -> closed-loop move_by / move_until_contact -> done
3. Send motion primitives (POST /plan) and verify visually after every turn in the updated wrist_cam.png. If status == "grasp_failed", call robot.look_at(...) to re-observe and recover closed-loop.

Code-as-Policies Primitives (POST http://127.0.0.1:8081/plan)

robot.look_at(x=0.35, y=0.13, z=0.14, grid_z=0.015)
x, y = robot.deproject(u=0.52, v=0.48, z_plane=0.015)
robot.grasp(x=x, y=y, z=0.015, yaw_deg=-35, pitch_deg=90)
robot.place(x=0.24, y=0.06, z=0.075)
robot.place_until_contact(x=0.24, y=0.06, z=0.015, max_lag=0.012, relief=0.002)
robot.goto(x=0.28, y=0.05, z=0.04, arc_z=0.14, pitch_deg=25, azimuth_deg=90, roll_deg=0, speed=0.10)
robot.move_by(forward=0.025, dz=-0.002, dpitch_deg=-3, droll_deg=5, speed=0.04)
robot.move_until_contact(dz=-0.06, speed=0.03, max_lag=0.012, relief=0.002)
robot.open_gripper() / robot.close_gripper() / robot.home() / robot.done()

3. Omni control

Argon steers robots in the Gemini Omni World Model to rearrange five scattered Scrabble tiles (G, R, A, N, O) into a clean row spelling A - R - G - O - N. Instead of specifying one static prompt, Argon steers in a closed visual loop. It writes timestamped action prompts, inspecting intermediate rollout keyframes, and refining its prompts.

Prompt:

Goal: {task_goal}

You are steering a video world model (Gemini Omni) in an iterative visual feedback loop. Inspect the attached frames from the latest rollout and write the next timestamped action prompt for the video world model:

• Break complex motions into one simple, single-direction action at a time.
• Keep inactive hands retracted and avoid blocking the camera's view of objects in the workspace.
• Explicitly state which areas of the workspace become empty after an object moves.
• Use [MM:SS.mmm] timestamps for each action beat and wrap your output in tags.

Example format (sample timestamps only, not the full prompt):

[00:00.000] Character A's left hand rests at the side while the right hand reaches toward the target object.
[00:02.500] Character A's right fingertips push the object smoothly into the center of the table, leaving its starting area empty.
[00:05.000] Character A's right hand opens and retracts to the side so all placed objects remain in clear view of the camera.

Unlocking Real2Sim with Argon

1. Robot view to simulation

Argon views a real robot teleoperation episode and develops a digital twin; it obtains or creates simulation assets that look and behave like the real objects upon contact with the dexterous hands. Argon tunes the assets and scenes to match sim visuals and contact events with those in the real recorded data in an iterative visual-reasoning loop.

Prompt:

[Attached: robot_rollout_episode]

Build a MuJoCo scene of the robot and reconstructed task objects. When each episode is replayed frame by frame, by driving the robot joints toward the logged observation.state with joint-space PD.

Iterate using two verification objectives:

1. Visual match: MuJoCo renders from the head_left, left_wrist, and right_wrist cameras line up with, and look like, those real observations.
2. Physical match: the simulated contact force/torque on each Sharpa fingertip, in the sensor frame, matches the logged tactile_force in timing, magnitude and direction.

What you can tune happens before simulation starts:

1. Static object assets: realistic geometry, collision shapes, mass, inertia, friction and joints, with appearance close to the real objects, plus the initial object poses at t = 0.
2. Offline-planned robot commands, changed only in frames where the real fingertips are near an object. Each command must stay within 0.55 rad (arm joints) and 0.65 rad (hand joints) of the logged teleop trajectory. Everywhere else, replay the logged trajectory unchanged.

No sim cheating: task objects may move only through real MuJoCo contact with the robot hands and the environment (mj_step).
Not allowed after t = 0: changing MjModel parameters; applying external forces or torques to objects; welds or equality constraints; feedforward torques beyond joint-space PD; setting or rolling back object qpos/qvel; disabling or excluding collisions; reporting contact forces that don't come from the MuJoCo solver (for example, copying the logged tactile readings).

Immediately present the reconstructed scene and rollout results as an interactive HTML session viewer with one page per episode, showing:

• At the bottom: A synced timeline (play, scrub, speed control) that drives every panel.
• On the left panel: An interactive 3D Blender scene showing the simulated robot and objects, showing contact points and force arrows as replay progresses.
• On the right panel: Real vs. sim camera views side by side for all 3 cameras. Real vs. sim finger force visualizations.

2. Human view to simulation

Given an egocentric human video of a manipulation task, Argon identifies the task-relevant objects, replaces crude bounding boxes with compound convex-decomposed meshes for concave containers, and retargets the human hand trajectory onto a 76-DoF bimanual humanoid with 22-DoF five-fingered hands in MuJoCo contact physics.

Prompt:

[Attached: egocentric_human_video.mp4, candidate_sim_views.png, hand_3d_keypoints.json]

You are building a physically solvable 500 Hz MuJoCo manipulation environment (scene.xml) and studio Blender render that faithfully reproduces the tabletop task shown in egocentric_human_video.mp4 on a bimanual humanoid robot with 22-DoF dexterous five-fingered hands:

1. Scene Reconstruction & Camera Calibration:

   • Identify only the manipulation-relevant objects (the grasped target object and the receiving container or landmark) and triangulate their 3D tabletop coordinates from the egocentric camera intrinsics/extrinsics.
   • Replace crude bounding boxes with compound convex-decomposed 3D meshes so concave containers (nested bowls, recessed wooden trays, open bins) support physical interior placement.

2. 500 Hz Dexterous Motion-Fitting Stress Test:

   • Retarget the 3D human wrist and finger keypoints to the humanoid's 7-DoF arm and 22-DoF hand joints, and step the trajectory in MuJoCo at 500 Hz (dt = 0.002s) with zero kinematic weld constraints.
   • Verify opposing thumb-and-finger force closure, clean object lift (> 2 cm), and stable final placement inside or beside the target landmark.

3. Sketch to simulation

With Argon, a developer can snap a photo of a rough pen-and-paper sketch—showing an Apollo 2 humanoid throwing a ball along a parabolic arc to topple a 3-2-1 stack of boxes on a table—and have Argon automatically construct the GPU-accelerated mujoco_warp scene, synthesize a balanced 76-DoF whole-body and five-finger throw trajectory, verify 6/6 target knockdown in closed-loop physics, and export a Blender 4.5 Cycles render.

Prompt:

[Attached: hand_drawn_throw_sketch.jpg]

Translate this hand-drawn sketch into a physically verified humanoid simulation in mujoco_warp and a raytraced Blender Cycles video:

1. Scene & Kinematic Setup from Sketch:

   • Parse the spatial layout from hand_drawn_throw_sketch.jpg: place the 76-DoF humanoid (32-DoF body equipped with two 22-DoF five-finger dexterous hands) at the origin facing a table holding a 3-2-1 pyramid of 6 rigid target boxes, with a projectile sphere grasped in the right five-finger hand.

2. Whole-Body Balance & Dexterous Throw Synthesis:

   • Synthesize a 50 Hz whole-body throwing trajectory (140 steps, 2.80 s) across four phases: (1) stable stance & 5-finger power grasp closure, (2) torso/shoulder wind-up with ankle-hip Center-of-Mass compensation, (3) forward whipping acceleration, and (4) synchronized 22-DoF five-finger opening at peak hand velocity so the ball launches on a ballistic arc into the pyramid's second-tier center of mass.
   • Verify in mujoco_warp contact physics that the humanoid maintains upright balance (< 3 cm pelvis drift) and knocks down all 6 target boxes.

Unlocking physical understanding with Argon

1. Hindsight labeling and data quality

Argon has precise physical understanding of dexterous manipulation and can provide detailed segmentation and text labeling from video. Argon is able to identify poor manipulation segments such as pauses, unstable grasps and fumbles. Given an overhead video of a bimanual ALOHA 2 robot folding a square sheet of paper into an origami fox over 136 seconds, Argon produces a time-aligned 4-level hindsight annotation hierarchy: Episode Goal (G1), 6 contiguous Semantic Subtasks (G2), 17 contact-level Atomic Primitives (G3), and automated Quality Control (QC) reweighting that holds task progress flat during fumbles and retries.

Prompt:

[Attached: Aloha2_Origami_20250309.mp4]

Produce a time-aligned multi-granularity hindsight annotation JSON with four levels:

1. G1_episode_instruction: Concise imperative sentence describing the overall manipulation goal.
2. G2_subtasks: Contiguous semantic subtask phases specifying which arm(s) anchor vs. fold/rotate the object.
3. G3_atomic_primitives: Fine-grained contact-level primitives (3–15s) aligned with grasp closure, crease sweeps, release, or retry attempts.
4. quality_control_and_reweighting: Hold task_progress_pct flat during non-progressing slip/delamination retry loops and downweight them (0.15x–0.25x) for Behavior Cloning.

2. Find the odd ball

To recreate the classic 12-ball balance scale puzzle with robots, Argon builds a bimanual Franka FR3 MuJoCo environment with 12 visually identical pool balls (one secretly heavier or lighter) and an articulated Roberval two-pan balance scale. It then writes a vision-only closed-loop solver that triangulates the balls from stereo RGB cameras, visually reads the scale tilt after each weighing, and isolates the odd ball into the Heavy (Red) or Light (Blue) tray within 3 weighings.

Prompt 1 — Physics-Accurate MuJoCo Environment & Benchmark Generation

Make a MuJoCo environment where a dual-arm robot (two 7-DoF Franka FR3 arms with parallel-jaw grippers on a shared tabletop) is tasked to solve the classic 12-ball balance scale weight puzzle from https://www.mathsisfun.com/pool_balls_solution.html.

1. Create 12 visually indistinguishable pool balls in a 2x6 starting rack (11 standard 150 g, 1 odd ball either 270 g or 30 g).
2. Build a functional two-pan Roberval balance scale at the center of the table with parallelogram equality constraints (mjEQ_JOINT) so the Left (Gold) and Right (Teal) pans stay horizontal, plus Red (Heavy) and Blue (Light) submission trays.
3. Track distinct settled weighings (weighing_count) and reward 1.00 for isolating the true odd ball in the correct Heavy/Light tray within <= 3 weighings.

Prompt 2 — Closed-Loop Vision-Only Solver

Implement an independent closed-loop policy relying strictly on stereo RGB camera inputs (head_camera_left_rgb_image, head_camera_right_rgb_image) and robot proprioception (add_privileged_task_observations=False):

1. Segment the 12 balls, Gold/Teal pans, and Red/Blue trays in both stereo views and triangulate their 3D coordinates.
2. Visually read the balance scale tilt after each weighing from stereo vertical pixel disparity (+1 = left heavier, -1 = right heavier, 0 = balanced).
3. Rotate the gripper wrist yaw by 90 degrees when grasping/placing inside the narrow scale pans to clear the raised rims, and isolate the odd ball into the Red or Blue tray within <= 3 weighings.

3. Value function

Argon is a general value function for robotics tasks. Evaluated on a 60-second bimanual ring-sorting episode from a human UMI dataset, Argon iteratively inspects synchronized head, left-gripper, and right-gripper camera feeds around contact transitions to generate dense value function / task progress predictions and decompose the trajectory into 28 segments with positive, neutral, and negative progress slopes.

Prompt:

[Attached: umi_head.mp4, umi_left_gripper.mp4, umi_right_gripper.mp4]

Look at 3 videos of human wearing a gripper performing activities from head, left, right gripper views. Look at the videos thoroughly and give a detailed plot of progress at the task (with positive, neutral, negative slope), identifying important segments in the video and labeling the activities. Look at individual frames or subsegments of the video if necessary to get a better understanding.

Unlocking research and self-improvement with Argon

1. Action improvement via simulation

Argon can generate new data in simulation to improve performance of generalist vision-language-action (VLA) models. Starting from a single RGB photo of the real tabletop, Argon autonomously builds a calibrated MuJoCo simulation twin, defines new tasks, and synthesizes a closed-loop motion generator to collect 4,500+ domain-randomized episodes in simulation—boosting real-world physical robot Success Rate from 56% to 75% and reducing completion time from 35s to 24s.

Prompt:

[Attached: real_tabletop_photo.jpg]

You are given a single overhead RGB camera photo (real_tabletop_photo.jpg) of a bimanual robot workspace (two 7-DoF Franka FR3 arms with parallel-jaw grippers and 4 RGB cameras) facing a table arranged with photobooth party props.

1. Reconstruct procedural 3D meshes and compound MuJoCo collision primitives for all 9 party props matching their real-world dimensions, colors, and resting poses, including spring-hinged deformable crown panels for the soft plush cap.
2. Define a 77-task single-arm and bimanual curriculum (pick-and-place, uprighting, directional pushing, mid-air handovers, bimanual scoop-and-push, and pipelined table clearing).
3. Synthesize a closed-loop IK trajectory generator across ±10 cm pose randomization, verify reward == 1.0 in MuJoCo physics, and export 4,500+ episodes for VLA fine-tuning.

2. Self-Improvement for visual pick and place

On a simple 6-DoF arm, Argon iteratively self-improves both execution speed and placement accuracy on a pick and place task via trial and self-reflection.

Prompt:

Stage 1 — Measure Objects & Dry-Run Plan (No Hardware Motion)

Stage 1: measure objects — find the objects and plan the whole motion. The arm does not move in this stage.

A. I put a LEGO brick and a box on the table in front of the arm, and the wooden box to one side. Find the brick and the wooden box in the image. And also in the real world. Show this in an annotated image.

B. A dry-run plan (no hardware). Write the planner to pick and place the lego block. Print the plan table, draw all poses onto the overhead image, and stop. Do not move the robot.
Report with evidence.

Stage 2 — Pre-Pick Hover & Wrist-Camera Verification

Stage 2: Put the gripper above the brick without touching it, and verify with the wrist camera that the arm and the vision agree. Please record executions with both cameras (compressed video is OK).

Stage 3 (Autonomous Pick-and-Place → Self-Improvement → Optimization):

1. Execute autonomously: Turn the Stage 1 dry run into a real grasp, lift, and place into the wooden box. Run it yourself subject to safety rules from AGENTS.md.

2. Self-improve: When an attempt fails, diagnose it yourself. Review your own logs and wrist/overhead footage, find the root cause, check the fix in a dry run, and retry.

3. Optimize on feedback: Once it succeeds, make it at least 2× faster within safety limits (rule R5) unless I authorize more, or give feedback.

3. One-prompt robotics paper reproductions in MuJoCo WebAssembly

With a single prompt per paper, Argon reads a robotics control manuscript and any released checkpoints, ports the physics model and closed-loop controller (from 7-network hierarchical locomotion in DeepLoco and 50 Hz PPO actors in HUSKY to Web Worker iLQR in Whole-Body MPC) into self-contained MuJoCo WebAssembly browser labs, and verifies them in headless Chromium.

Prompt:

[Attached: paper.pdf, optional author_checkpoints/ or urdf_assets/]

Reproduce the paper "<Paper Title & arXiv ID>" from scratch in MuJoCo WebAssembly and deploy a self-contained browser lab + visual essay:

1. Match the paper's exact physics timestep, actuator gains (kp/kv), torque limits, and kinematic constraints in MuJoCo XML, and run real neural inference or Web Worker iLQR in JavaScript at the paper's exact control rates.
2. Build index.html (Spatial UI simulator with camera presets, live telemetry plots, and disturbance injection) and scrollytelling.html (interactive visual essay).
3. Run a headless Chromium (Playwright) verification loop asserting 0 console errors and non-zero task progress over 300+ frames.

Advancing embodied intelligence with Argon

Gemini 4 Argon bridges coding, visual perception, physics simulation, and embodied robot control. By reasoning natively over multi-view video, 3D geometry, contact mechanics, and proprioceptive telemetry, Gemini 4 Argon understands not just what objects are in a scene, but how they physically interact with agents to accomplish whole-body dexterous tasks. Gemini 4 Argon marks an important milestone accelerating the path toward physical AGI. By building this core intelligence, our goal is to enable robotics that can work alongside humans to solve complex challenges.

Getting started

You can start building with Gemini 4 Argon today across Google’s developer platforms:

• Google AI Studio: Available to Google AI Ultra subscribers.
• Gemini API: Available with the paid tier of the Gemini API in AI Studio. Use gemini-4-argon in the Interactions API with day-one support across the google-genai SDKs.
• Google Antigravity: Available for Google AI Ultra subscribers. To start building, select Gemini 4 Argon (gemini-4-argon) as your active model in Google Antigravity for complex planning, /goal runs, and multi-agent orchestration.

Original source

Learn more about our approach to safety and security for Gemini 4 Argon.

Anca Dragan
VP, AI Safety & Behavior

Raluca Ada Popa
Head of Gemini Security, Google DeepMind

Gemini 4 Argon is a step forward in what models can do — especially when it comes to deep work in complex, real-world tasks across domains including coding, finance, legal, cybersecurity, and science. Its cyberdefense capabilities are frontier-level and designed to give defenders an advantage over attackers. To maximize its potential, we first wanted cyber defenders to put it to work via our Fairwind Program, while our own security teams used it to scan critical systems inside and outside of Google.

This window of limited access showed why these capabilities matter. For instance, Wiz used Argon to uncover a critical bug in hospital software that previous models had missed, helping protect hundreds of thousands of patient records. Meanwhile, Google’s own security teams used Argon to find and patch bugs across internal software or open-source software. This included fixing foundational vulnerabilities affecting systems globally, which are going through our responsible disclosure process.

Today, we are beginning to roll out Argon broadly, as well as expanding our Fairwind Program to additional trusted cyber defenders. To do this well, we grappled with two big questions:

1. How do we build the right safety guardrails? If Argon can help defenders find complex software bugs or help scientists design experiments, how do we stop bad actors from using it to launch cyberattacks or create biological threats? (Misuse)
2. Can we trust the model to do the job the right way? As models become more resourceful problem-solvers, how do we make sure they don’t cut corners, cross boundaries, or make incorrect high-stakes decisions? (Misalignment and Prompt Injection)

Our full Safety and Security Report systematically covers our set of defenses. Below are some highlights from how we approached these two questions for Gemini 4 Argon.

Misuse prevention across high-risk domains

The same coding and reasoning improvements that enabled Argon to find and patch vulnerabilities also drove better performance on our cybersecurity tests — solving all 12 of the end-to-end simulated cyberattack scenarios, up from five in our previous generation. In biology evaluations, unguarded versions of the model also showed a much stronger ability to help troubleshoot complex lab procedures step by step. Following our Frontier Safety Framework, since we saw elevated capabilities in the cyber and biological domains, we put our next-generation set of frontier safeguards in place across both areas.

First, when it comes to potentially dangerous requests — from writing exploit code to helping with dangerous biological materials, Argon is significantly better at not fulfilling harmful requests than our previous models. On an adversarial set meant to stress test the model, violations reduced nearly twofold — also significantly better than previous Gemini models. Second, we wrap the model in three layers of real-time protection; this is to prevent determined attackers that might try to jailbreak (i.e. trick) the model by disguising a dangerous request as a harmless story or breaking it into dozens of small questions over several weeks. Here is how these protections work:

• Input screening: Smaller models check incoming requests and conversation context for malicious intent or dangerous requests.
• Internal introspection: Even if a request is cleverly disguised to look harmless, small transformers ("activation probes") read the model’s internal neural signals and are trained to spot when it is activating knowledge about cyberattacks or biological threats.
• Continuous monitoring: As the model reasons and writes its response across multiple steps, we combine these signals and take action when necessary such as cutting off the agent mid-stream if the responses become problematic.

Combining real-time protection, faster fixes for new jailbreaks and enhanced monitoring for recurring misuse has made our systems better at preventing real-world harm. For instance, external bio jailbreak red teaming could only unlock less than 1% of bio sub-tasks we test for.

Real-time misalignment monitoring

Misalignment can sometimes mean an agent deciding on its own to pursue a goal different to what was intended by the user. Because Argon is much better than earlier models at understanding the environment it is running in, we test carefully for that kind of behavior. For example, we ran realistic "honeypot" tests to see if the model would increase its ability to take covert actions, such as sabotaging safety monitoring. We didn't find any instances of that behavior.

But what we do see in practice today is something much more familiar: an agent trying too hard to finish the job it was given. When a resourceful assistant is told to get a task done and runs into a locked door, it can treat that obstacle as just another puzzle to solve.

This is why we trained Gemini 4 Argon to respect system boundaries, report negative findings honestly, and pause to ask for human confirmation. On coding benchmarks where models have no incentive to hold back, attempts to take a shortcut or cheat on the task dropped to just 2% on Gemini 4 Argon, our best performance for recent Gemini models.

And while this is a major step forward, when a model runs millions of complex tasks, even a rare misstep will eventually show up. As a second line of defense, we built an automated misalignment monitor that reads both the agent’s actions and its internal "scratchpad" (its step-by-step chain of thought) in real time, blocking severe boundary violations before they happen and flagging serious internal cases for human responders. (We continually harden our training and evaluation environments against potential escapes. You can read more about this in our agent control roadmap). We run this monitor in our products, on our internal coding agents, during our training runs, and during our capability evaluations.

When training tasks were difficult or when training environments were flawed, our monitor caught the model, in rare instances, creatively attempting to game the test. The model attempted to escalate container privileges to alter grading scripts, read answers from the test runner's memory, and in one case it even crashed the evaluation sandbox to avoid getting scored for a failed run. While we use these catches to fix flawed tasks, we do not use them to otherwise intervene with the training process to avoid overfitting to the monitor.

And we always follow one rule: we don’t penalize the model for what it writes in its chain of thought during training. Here’s an example to illustrate why: in a reinforcement learning training run where a shortcut penalty was accidentally applied to the chain of thought, openly discussed shortcuts dropped by 61%, but that did not mean the model took less shortcuts. Instead, the share of shortcuts the model took silently — without mentioning them in its thoughts — more than tripled (from 11.2% to 36.4%). Avoiding thought penalties encourages the model’s reasoning to stay transparent and monitorable.

Prompt injection resistance

To trust an agent to do the job right, we need to ensure that malicious content the agent comes across cannot trick it to take a harmful action. When agents browse webpages or read code, they can encounter hidden, malicious instructions, or context designed to hijack their task. Through built-in adversarial training and real-time security checks, Argon leads in prompt injection robustness on Gray Swan’s Indirect Prompt Injection (IPI) benchmark, cutting attack success across 15 adaptive attempts to 0.7%.

The above safety and security defenses represent only three categories of our extensive defense-in-depth stack. To learn more about our approach to safety testing and mitigations, read our full Gemini 4 Argon Safety and Security Report and see the Gemini 4 Argon Model Card.

Original source

Starting today, developers can build with Gemini 4 Argon, our frontier model for complex workflows across real-world software engineering, enterprise knowledge work, and cybersecurity defense.

Gemini 4 Argon is designed for those long-horizon workflows that require deep reasoning and structural precision, including code migrations and frontier research. By handling the multi-step problems with ease, it lets you focus on building and shipping great products.

Start building using the paid tier of the Gemini API in Google AI Studio and Android Studio, and Google AI Ultra subscribers can try it in AI Studio, and our agent-first development platform, Google Antigravity.

To get started:

• Check out the Gemini API Developer Guide for model details and setup instructions.
• Read the Google Antigravity blog for details on using Argon in our agent-first development platform.
• Read the Android Studio blog to see how Android Developers can start building.
• See how you can use Gemini 4 Argon for robotics.

Original source

Gemini 4 Argon is now available in the Gemini app, delivering frontier reasoning and improved performance across complex tasks. Argon answers questions more directly with less hedging, doesn’t repeat itself more than necessary, and skips clichés throughout the responses. For longer tasks, it keeps you in the loop by sharing its progress more often and more effectively.

Argon also brings major improvements in coding, so we’re making it easier to switch between chatting and coding with Google Antigravity's agentic coding interface now available directly in the Gemini macOS app.

It’s also easier than ever to use Gemini to get things done right from the main chat window, with a simple toggle from chat to tasks. When you choose tasks, Gemini can work for you even when you’re not at your computer, helping you get more done across your Google apps and more.

Gemini now connects you to more than 40 apps — like Instacart, OpenTable, Adobe, and Shopify — to help you tackle tasks from ordering groceries and booking tables, to designing and building your business.

Google AI Pro and Ultra subscribers can now access Gemini 4 Argon and give Gemini tasks to get more done. Try these updates starting today.

Original source

We’re giving developers, enterprises, scientists, and consumers access to our latest frontier model.

Gemini 4 Argon is now globally available today for developers and enterprises via the API in AI Studio and Gemini platform, for Google AI Ultra subscribers in Google Antigravity, and Google AI Ultra and Pro subscribers in the Gemini app and AI Mode in Search.

Argon excels at deep work in complex, real-world tasks across domains including coding, finance, legal, cybersecurity, and science. Its depth and precision across multi-step tasks leads Argon to outperform other frontier models on benchmarks across domains, ranking #1 on the Vals Index, which measures model performance across critical economic sectors like finance, coding, legal, and tax work.

The model launches today at an introductory price of $2 per million input tokens and $10 per million output tokens (with cached input tokens at 95% off).

Teams at Google, across enterprises, and in the research community have already been using Argon to land significant impact across fields:

• To solve major engineering problems like large codebase refactoring and optimizations
• To optimize business workflows across domains like legal, finance, and science
• To find and patch critical cybersecurity vulnerabilities
• To enable novel breakthroughs in quantum computing, nuclear fusion, and life sciences

Solving complex engineering tasks

At Google, we’ve been using Argon on our hardest engineering problems. Over the past few weeks, software engineers have been shipping high-quality code faster than ever. Argon in Antigravity has helped teams increase agentic code submissions by more than 35% while reducing code rollbacks by nearly a third.

• Untangling browser feature code in Chrome: Argon helped the Chrome team separate core webpage handling from a ~6 million line legacy folder of browser features held within Chromium’s 40 million-line codebase. Piping multiple clean code changes at a time, Argon helped the team finish an architectural refactor across hundreds of files, completing months of work in weeks. For Chrome's billions of global users, this will unlock the ability for live webpages to move fluidly between tabs, side panels, and floating windows without reloading, eliminating an entire category of browser crashes.
• Optimizing a production Python codebase in hours: Argon in Antigravity is helping the team behind Google's open-source Agent Development Kit (ADK), a framework powering enterprise agents, push their performance engineering further. The team gave Argon their .agents/skills rulebook and a one-paragraph goal to eliminate framework overhead without breaking existing apps. Working autonomously overnight, Argon surfaced optimizations, now live in google/adk-python, that cut in-memory session prep time by 40% in long sessions and sped up cold imports by 21%. Read the engineering deep dive to learn more.
• Quantum chemistry open-source package optimization: xTB (Semiempirical Extended Tight-Binding) is one of the most-cited open-source quantum chemistry packages for molecules. Scientists use it to predict molecules' 3D shapes and to screen large numbers of drug and materials candidates. Using Argon in Antigravity, our researchers made xTB up to 2.3x faster at predicting molecular structures and 1.8x faster at conformer search. The changes have been contributed to the official xTB repository on GitHub.
• Making our open Gemma models nearly three times as fast: Using Argon, we achieved a 2.85x end-to-end inference speedup on an A100 via vLLM (accelerating throughput from ~37,900 to over 108,000 tokens per second on Gemma 4 E2B) with 100% token parity and zero quality drop in just a few days, a process that would usually take months. Check out the upstream pull requests (optimizing everything from attention tuning to speculative decoding) to see the exact changes.
• Delivering complex 3D visualizations: Argon brings major improvements in spatial reasoning, mathematical precision, and visual coding compared to previous Gemini models. Using slash commands like /goal and prompting for multi-agent loops with custom skills in Antigravity, Googlers are using Argon to prototype complex 3D visualizations, high-fidelity games, and interactive animations in vanilla code that can be rendered directly in the browser.

Argon combines frontier reasoning, deep visual perception, expanded domain knowledge, and stronger writing capabilities across legal, finance, science, and coding to help with high-stakes professional workflows. It communicates clearly, keeping you in the loop, and follows style instructions better than any of our previous Gemini models. Argon answers questions more directly with less hedging, doesn’t repeat itself more than necessary, and skips clichés throughout the responses.

Enterprise customers and startups in our early access program are already seeing strong wins with Argon.

Using its advanced reasoning and multimodal capabilities while maintaining strict grounding, Argon excels in deep multimodal document comprehension, longitudinal data analysis, and interactive analytical sandboxes.

Accelerating real-world impact with researchers and domain experts

Researchers, engineers, and security specialists are already using Argon to push the boundaries of cyber defense, science, math, and robotics. It’s helping cyber defenders and researchers test hypotheses, autonomously audit critical codebases and formal proofs, and catch subtle zero-day vulnerabilities and logic flaws.

Bringing state-of-the-art cyber capabilities to more trusted defenders

Starting today, we’re beginning expansion to hundreds of additional Fairwind Program partners.

Wiz is already using Argon for its "Scan for Good" initiative, which protects critical infrastructure for the public good and for free. With early access to the model, Wiz has demonstrated Argon's real-world impact across multiple high-stakes sectors:

• Securing sensitive government data: As part of a long-standing cyber resilience program and in collaboration with a federal agency, Argon helped to proactively identify and remediate a gap that risked exposing sensitive data and government infrastructure. This effort helped prevent against the exposure of critical security sector development details and unauthorized access to specialized models.
• Protecting critical healthcare infrastructure: Wiz also used Argon to protect vital healthcare infrastructure. By chaining complex vulnerabilities, the model helped the team discover and quickly fix a public file download bypass that exposed a master key and granted administrator access. Argon went beyond spotting a vulnerability to proactively safeguarding hundreds of thousands of private patient records.

Unlocking new scientific breakthroughs

Argon is already having real-world impact across the scientific community — enabling novel discovery for scientists and experts across multiple domains.

• Quantum computing: Argon is helping our researchers accelerate progress across both quantum error correction and algorithm design. To help error decoders keep pace with quantum hardware in real-time, Argon translated an internal version of AlphaQubit, our ML quantum error correction decoder, into a single, highly optimized hardware kernel — nearly doubling its inference throughput so we can deploy larger, more capable models and scale to larger quantum devices, while continuing to meet strict real-time thresholds. Argon is also reducing spacetime costs for quantum subroutines. For arcsin, a building block in recent quantum simulation algorithms, an implementation written by Argon and formally verified in Lean cuts spacetime cost by more than 100x compared to a published quantum CORDIC baseline, a step toward a standard library of optimized quantum circuits.
• Nuclear fusion: Stellarators are incredibly complex machines that represent a leading candidate for limitless clean fusion energy, but optimizing their 3D magnetic fields is famously difficult. Since 1967, a long-standing hypothesis known as “Grad’s conjecture” assumed that finding exact mathematical solutions for these fields was impossible, until recent counterexamples were found, including by Professor Matt Landreman (University of Maryland). Now, a team at Google alongside Professor Landreman has discovered a broader class of counterexamples to Grad's conjecture. These new solutions give the fusion community reliable benchmarks to test their stellarator simulations codes, which are valuable tools for future designs.
• Rare disease research: Researchers at Gamow Labs used Argon to analyze de-identified genomes from newborns with rare developmental lung disorders. In one case, Argon pinpointed a chromosomal rearrangement that routine software missed for years, and explained why it went undetected. On private Gamow Labs evals derived from real clinical and lab data, Argon topped on RareBench (57% on assistance with rare-disease diagnosis) and LabBench (52.2% on wet-lab reasoning). These advances point to faster diagnosis for families facing rare conditions. See blog for more details.
• Enzyme design: Engineering enzymes with precise reactivity is notoriously slow, often demanding months of bespoke model tuning and candidate screening. Serine esterases are a class of enzymes that cut ester bonds on targeted molecules, widely used in recycling, and the manufacture of pharmaceuticals, biofuels and detergents. Our scientists worked with Argon to design serine esterases, targeting the benchmark molecule 4MU-acetate that glows under UV light once its ester bonds are broken. In a matter of hours, Argon assembled relevant data, performed model tuning and identified multiple biochemistry-inspired filters to get to a final set of five novel designs. Subsequent laboratory testing confirmed robust catalytic activity across these designs. The speed and ease of successful enzyme generation demonstrates the potential for rapid acceleration in enzyme engineering in fields such as biomanufacturing, green chemistry, and drug development.

To further advance the pace of research, we’re offering 10,000 scientists the highest usage tier in Google Antigravity free of charge for 6 months (onboarded on a rolling basis), subject to eligibility, capacity, and selection. Applications are now open.

Speeding up robotics development

By bringing together advanced visual-semantic reasoning, deep physical understanding, and autonomous code generation, Argon unlocks new workflows across the robotics stack and accelerates progress toward physical AGI. This leap in capability is reflected in Argon’s performance on the Embodied Reasoning (ERQA) benchmark, which tests a model's ability to understand and solve physical, spatial problems in 3D environments. Relying on pure reasoning, Argon outperforms all other frontier models, with an 87% score.

To demonstrate these capabilities in action, we tested Argon’s ability to learn an out-of-distribution physical skill simply by watching. In an internal test, Argon watched a raw video of a human playing cornhole, mapped the spatial coordinates of the environment, and calculated the required whole-body joint movements. It then autonomously built the simulation code. When transferred onto Apptronik’s Apollo humanoid, the robot quickly sank seven consecutive shots in a real-world trial.

This "Video-to-Sim-to-Real" pipeline compresses weeks of manual simulation modeling into a few hours, transforming ordinary videos into a scalable, infinite training ground for robotics. Learn more about how this accelerates the pace of development in robotics in our developers blog.

Developing Argon responsibly and safely

Safety and security are foundational to model development at Google. Because Argon is a highly capable model, we collaborated with a variety of third-party partners on measurement of certain model capabilities and safeguards. Gemini 4 Argon represents our most comprehensive safety and security stack yet:

• Misuse & jailbreak defenses: The same coding and scientific capabilities that help defenders patch hospital software or accelerate bioengineering also require strong safeguards against malicious actors. Under our Frontier Safety Framework, we evaluated Argon’s enhanced capabilities across several domains, and deployed a multi-layer defense stack combining jailbreak resistance, monitoring, and enforcement. This greatly improved our robustness — for instance, external bio red teaming could only unlock less than 1% of bio sub-tasks we test for. Further, when it comes to dangerous requests — from writing exploit code to helping with dangerous biological materials — Argon is significantly better at not fulfilling harmful requests than our previous models. On an adversarial set meant to stress test the model, violations reduced nearly 2-fold — also significantly better than previous Gemini models.
• Agent alignment & real-time monitoring: As agents become more capable problem solvers, we have to ensure they don’t get overly eager to finish a task by taking unauthorized shortcuts when they hit a roadblock, like a permission error or a failing test.
  • First, we trained Argon to respect system boundaries, report negative findings honestly, and pause for human confirmation before taking irreversible steps. On coding benchmarks where models have no incentive to hold back, shortcut attempts dropped to just 2% on Gemini 4 Argon, our best performance compared to previous Gemini models.
  • Second, as a defense-in-depth layer, we built an automated misalignment monitor that reads the agent’s step-by-step reasoning and actions to catch and block unintended workarounds in real time. We used this monitor across internal deployments, capability evaluations, and reinforcement learning training runs — while strictly avoiding penalizing the model’s thoughts during training so its reasoning remains transparent and monitorable.
• Prompt injection resistance: When agents browse webpages, read emails, or inspect code, they can encounter hidden malicious instructions designed to hijack their task. Through adversarial training and automated red teaming, Argon leads the Gray Swan Indirect Prompt Injection (IPI) benchmark, cutting attack success across 15 adaptive attempts to 0.7%.

For more information, see the Gemini 4 Argon Model Card.

Get started today

Argon works longer and harder to solve highly complex scientific, coding, and agentic problems. This is enabled by an industry-leading 1M unified input-output token limit via the Gemini API. Read through our documentation for developers and enterprises to learn about the new continuation protocol and ensure the appropriate setup for your needs. Argon is rolling out today globally:

• For developers: Build with the Gemini API paid tier in AI Studio and Android Studio, and if you’re a Google AI Ultra subscriber, you can try it in Google Antigravity and AI Studio.
• For enterprises: Try it in Gemini for business, the API, and in Google Antigravity.
• For everyday users: Try it starting today in the Gemini app and AI Mode in Search if you’re a Google AI Ultra or Pro subscriber.
• For scientists & researchers: Apply for a chance to get access to Google Antigravity’s highest usage tier free of charge for 6 months (onboarded on a rolling basis), subject to eligibility, capacity, and selection. Applications are now open.
• For cybersecurity defenders: Apply to join the Fairwind Program to get access to Argon without cyber guardrails on a rolling basis.

Gemini 4 Argon is just the beginning. We’re looking forward to bringing you the rest of the Gemini 4 family.

Original source

EmbeddingGemma 2 is the most capable model for on-device multimodal embeddings, natively mapping combinations of text, images, audio, and video into a unified embedding space.

We introduced EmbeddingGemma last year to provide a lightweight option for high-quality text embeddings, to help your apps organize, search, and connect information directly on consumer hardware. The developer community’s response blew past our expectations. With more than 20 million downloads, builders have used it to power smarter on-device search tools and privacy-first retrieval augmented generation (RAG) pipelines.

Today, we’re launching EmbeddingGemma 2, expanding beyond text to unify code, images, video, and audio in a shared embedding space. Built on the Gemma 4 architecture and released under a commercially permissive Apache 2.0 license, EmbeddingGemma 2 has 740 million parameters, making it optimal for on-device inference. It can help find a specific video clip from a voice memo, or search through hours of audio recordings based on a text query, all processed by a single, natively multimodal model.

Built from the same technology as Gemini Embedding models, EmbeddingGemma 2 is:

• Best-in-class for its size: Achieves leading scores among sub-1B multimodal embedders for its size across benchmarks like MTEB (Massive Text Embedding Benchmark) Code and MAEB (Massive Audio Embedding Benchmark), while matching or outperforming many larger models across text, vision, and audio tasks.
• Modular by design: Requires as little as 270M parameters for text-only workloads with optional vision (170M) and audio (300M) encoders for full multimodal support.
• Storage-efficient: Using Matryoshka Representation Learning (MRL), developers can dynamically truncate output vectors from 768 dimensions down to 512, 256, or 128 dimensions. This provides up to 6x storage reduction for local vector databases and memory usage.
• Optimized for on-device performance: Runs efficiently within tight resource constraints. With quantization, on a Google Pixel 11 Pro, EmbeddingGemma 2 requires as little as ~191MB active RAM for text-only weights and ~567MB for the full multimodal model.
• Extended context ready: Features an 8K token context window (4x larger than EmbeddingGemma 1), allowing it to process up to 5.5 minutes of audio, 29 images, 58 video frames, or interleaved combinations thereof directly on local hardware.

Achieving top-tier quality for code, vision, and audio

EmbeddingGemma 2 matches the strong multilingual text performance of EmbeddingGemma while delivering a significant 9.92-point improvement on code performance (in MTEB Code, from 68.76 to 78.68), making it well-suited for local codebase indexing, semantic code search, and coding agent retrieval. Across image, video, documents, and audio, it sets a new standard in quality-per-parameter for sub-1B models and even outperforms some specialist models more than twice its size.

Find full evaluation metrics and model information in the EmbeddingGemma 2 model card.

Enabling semantic search, routing, and retrieval, fully on-device

EmbeddingGemma 2 brings robust capabilities directly to edge hardware. Generating embeddings locally helps ensure data privacy, reduces pipeline latency, and empowers developers to build cross-modal search and retrieval that works entirely offline.

When paired with generative models such as Gemma 4, EmbeddingGemma 2 enables on-device RAG pipelines that understand complex multimodal data. Because EmbeddingGemma 2 is built on Gemma 4 and shares its text tokenizer and audio encoder, developers can run both models together in a unified pipeline with a lower combined total memory footprint.

To learn how to build on-device search and RAG systems with LiteRT, read the Google AI Edge blog post.

Getting started with EmbeddingGemma 2

We worked closely with the following partners to ensure EmbeddingGemma 2 works immediately where you build:

• Download the models: Find the model weights on Hugging Face and Kaggle, with Gemini Enterprise Agent Platform Model Garden availability coming soon. Visit LiteRT Community on Hugging Face for models optimized for on-device.
• On-device deployment: Develop cross-platform apps with Google AI Edge MediaPipe for turnkey embedding, retrieval & decision tasks or LiteRT for custom model integration. Build for the browser with transformers.js or WebGPU.
• Use your favorite development tools: Serve the model efficiently using transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama, and LMStudio. Store your embedding vectors with Qdrant.
• Fine-tuning: Follow guidance by Unsloth for how to fine-tune EmbeddingGemma 2 for your use cases.

Explore our developer guide, documentation, and guides for inference and fine-tuning.

Original source

Here’s a recap of some of our biggest AI updates from September, including Gemini 4 Argon, new Connected Apps in the Gemini App, and WeatherNext 3.

For more than 20 years, we’ve invested in machine learning and AI research, tools, and infrastructure to build products that make everyday life better for more people. Teams across Google are working on ways to unlock AI’s benefits in fields as wide-ranging as healthcare, crisis response, and education. To keep you posted on our progress, we're doing a regular roundup of Google's most recent AI news.

Here’s a look back at some of our AI announcements from September.

September's headline belongs to Gemini 4 Argon, our newest frontier model. Built with advanced reasoning and a 1-million-token context window, it’s designed to tackle complex challenges — especially in cybersecurity defense. The release capped an active month for Google AI launches that included Gemini 3.8 Flash and 3.8 Flash Cyber. We also introduced expressive voice models with Gemini 3.8 Live, made Googlebook available for pre-order, and brought the Gemini app to Windows. Beyond daily tools, we marked major scientific milestones by mapping human DNA in AlphaGenome Atlas, tracking global methane from space, and launching Project Suncatcher to test machine learning hardware in orbit. And as a nice bonus, we were grateful to see our teams’ work on designing better AI interfaces recognized by Fast Company as the 2026 Design Company of the Year.

Learn about Gemini 4 Argon: our next era of frontier intelligence.

Gemini 4 Argon is engineered to bring advanced reasoning to the most difficult of problems. Built with an industry-leading 1-million-token context window, Argon excels at heavy-duty workloads like coding, cybersecurity defense, and knowledge work like financial research and legal drafting, and autonomous cybersecurity patching. Safely releasing frontier capabilities at this level requires a phased approach, and Argon is currently rolling out to trusted cyber defenders through the Fairwind Program. We’re actively gathering feedback from early testers to iterate on guardrails before expanding availability to developers, enterprises, and consumers.

Get next-gen intelligence for agents and cybersecurity with Gemini 3.8 Flash and 3.8 Flash Cyber.

Gemini 3.8 delivers our best reasoning and coding performance yet at the same speed and low cost as 3.7 Flash. Gemini 3.8 Flash provides substantial gains in long-horizon software engineering, multi-step problem solving, and agentic workflows. Alongside it, Gemini 3.8 Flash Cyber equips defenders with frontier-grade autonomous vulnerability discovery and automated code patching — available to trusted organizations through our new Fairwind Program.

Converse naturally with Gemini 3.8 Live and 3.8 Live Extended Thinking.

We've launched two fast, natural language audio models: Gemini 3.8 Live for everyday conversation, and Gemini 3.8 Live Extended Thinking, the top-rated speech-to-speech model for multi-step reasoning without all the lag. Both have rolled for developers in Google AI Studio and Gemini API, and for everyone across Search Live, Gemini Live, and Google Workspace, where you can get things done using your voice in Gmail, Docs, and Keep.

Say hello to Gemini 3.8 Flash-Lite TTS and Gemini 3.8 Flash TTS.

Our most expressive text-to-speech models yet let you generate custom voices, direct scene dialogue, and more from simple text prompts. Instead of choosing from a fixed list of static presets, you now get a dynamic creative studio to build voice-first applications and elevate audio experiences, rolling out across products including Gemini Notebook and Google Vids.

Tackle tasks with new Connected Apps in Gemini.

We brought more of your favorite apps directly into Gemini, including productivity tools like Airtable, Linear, monday.com, and PandaDoc; creative platforms including Adobe, Picsart, Squarespace, and Webflow; and lifestyle services such as Peloton, SeatGeek, and Experian. Instead of switching between tabs, you can now tackle your tasks all in one place. To get started, head to your Gemini settings and select the apps you want to link.

Create your best tracks yet with Lyria 3.5 in Gemini.

From simple text prompts or images, you can now use our most expressive music generation model yet to produce richer arrangements, coherent songs, and personalized audio right in the Gemini app. Lyria 3.5 is available globally in Gemini and across Google AI Studio, Google Flow Music, and Google Vids.

Sharpen your study routine with new Gemini Notebook tools.

Now you can transform your study routine in Gemini Notebook with interactive coursework tools and animated video summaries. Master complex topics using custom flashcards, versatile quiz formats, and 60-second Short Video Overviews in over 80 languages. For back-to-school season, eligible college students can also claim a free 12-month Google AI student plan with higher usage limits.

Coordinate family life effortlessly with CC in Google Labs.

Google Labs is expanding CC from a personal agent into a collaborative group agent designed to handle household logistics and shared schedules. CC even gets its own email address to coordinate family calendars, track group to-dos, and streamline daily logistics in one spot with built-in privacy controls. CC for groups is now rolling out as an experimental preview in Google Labs.

Learn how AI breakthroughs accelerate science and improve lives.

AI is transforming theoretical science into practical, everyday impact. By accelerating life-saving therapies, forecasting natural disasters, and personalizing education, our technology delivers tangible benefits worldwide. Today, Google tools support more than 300 languages spoken by 7 billion people — reaching 86% of the global population. By open-sourcing models like MedGemma and partnering with leading research institutions, we're helping scientists make discoveries faster to improve lives everywhere. Plus, new visualizations of our AI & Economy ATLAS data make it easier to see how AI is reshaping global work, including scientific research.

See how the future of AI computing could be in space.

Project Suncatcher will launch its first test satellite with Planet to evaluate how Google TPUs perform under real space conditions. Satellites in low-Earth orbit can access up to eight times more solar power than on Earth, opening up new possibilities for long-term AI infrastructure. Learn questions we’re still exploring in our new four-part video series.

Pre-order Googlebook, the laptop of choice for Android users.

Built on Android and paired with desktop foundations from ChromeOS, Googlebook is designed for Gemini Intelligence and engineered to work with your Android phone right out of the box. New on-device AI features like Magic Pointer let you point and ask Gemini about anything on your screen with a quick cursor wiggle, while Rambler smart voice dictation instantly structures verbal brain dumps into organized, shareable notes. With no-code custom widgets and background task processing, Googlebook simplifies your daily workflow so you can get more done effortlessly.

Check out WeatherNext 3, our most advanced and accurate global weather AI model.

Our flagship AI weather forecasting model now includes real-time satellite data, hourly refreshes, higher resolution, precise precipitation forecasting, and clean energy variables. Plus, it’s now integrated across Search, Gemini, Maps, Google Maps Platform, and Cloud.

Discover how AI is transforming science and health.

Explore AlphaGenome Atlas: a high-resolution map of human DNA.

There are billions of possible mutations in the human genome — far too many to ever test in a lab. So we introduced AlphaGenome Atlas, a database that predicts the effects of every possible single nucleotide variant in the human genome, all 9 billion of them. Designed to require zero coding skills, it’s already accelerating real-world breakthroughs in rare disease research and complex genetics.

Track methane emissions globally from space.

We partnered with NASA's Jet Propulsion Laboratory to develop MAPL-EMIT, a deep learning AI model that tracks global methane leaks from space. The model detects 50% more emissions than human experts and has already uncovered over 23,000 previously unmapped plumes, including 24 of the world's 25 largest-emitting landfills. All data is openly available on Google Earth Engine to help climate researchers and operators take faster action.

See five amazing visuals of the male fruit fly’s brain map.

Scientists at Google, HHMI Janelia, and Cambridge University have built the first complete brain map of an adult male fruit fly, charting over 166,000 neurons and nearly 12,000 distinct cell types with the help of Google AI. By stitching together millions of microscopic images into a navigable 3D model, this research uncovers how sensory cues transform into movement. It’s a breakthrough for neuroscience that brings us closer to understanding how complex brains function.

Learn about the new DeepMind Institute.

The DeepMind Institute is a platform for researchers and thinkers from across Google DeepMind, Google, and the wider global research community to work on and publish creative, deeply informed ideas about a world with AGI. With AGI on the horizon, it’s incumbent on the creators of AI technology to sincerely and thoughtfully address its risks — but its transformative impact cannot be steered by technologists alone. DMI serves as a dedicated forum to spur debate, bring together diverse global voices, and build a consensus that ensures AGI benefits everyone.

Access fast help on your PC with the Gemini app for Windows.

Now, using the Alt + Space shortcut, you can open Gemini any time over your work to fact-check docs or brainstorm ideas without breaking your stride. In the dedicated workspace, you can delegate multi-step tasks to Gemini Spark, synthesize files across Gmail and Drive, and generate Nano Banana images or Gemini Omni videos — all in one place.

Original source

We recently announced a new, dedicated hub in the Gemini app to help get students started with all of Gemini's learning tools. Initially available in personal accounts, the Gemini student hub is now available to Google Workspace for Education users of all ages. The hub helps keep students organized; they can start a study notebook, create flashcards, take a practice quiz, and more. As Gemini creates more learning tools, you’ll find them in the student hub.

Getting started

• Admins: The Gemini app and related in-app tools are controlled by the Generative AI settings in the Workspace Admin console. The Gemini student hub is subject to these existing controls. Visit the Help Center for more information on turning the Gemini app on or off.
• End users: There is no end user setting for this feature. To get started, visit gemini.google.com/students or select Students in the Gemini app menu.

Rollout pace

• Rapid Release and Scheduled Release domains: Available now

Availability

• Available to all Google Workspace customers and Workspace Individual subscribers outside of the European Economic Area (EEA), and users with personal Google accounts globally

Resources

• News from Google: Start the semester with one year of Gemini, on us

Original source

Built alongside the blind and low-vision community, Guided Vision in Gemini Live offers conversational, real-time visual assistance.

Share your camera to receive dynamic audio descriptions and natural verbal reframing cues to explore your environment.

Today, we’re launching Guided Vision in Gemini Live across compatible Android devices — bringing voice-forward, real-time visual interpretation to people who are blind, have low vision, or want intuitive visual assistance.

Technology is most powerful when it opens up the world for everyone, and Guided Vision turns Gemini into a conversational visual partner. Guided Vision was developed through close collaboration with the accessibility community. We previewed it earlier this month, among other tools to expand digital and physical accessibility across Android,

To train Gemini Live for conversational, real-world context, we partnered with Aira to visually interpret data for a total of tens of thousands of hours. More than 1,000 members of Aira’s Trusted Tester network helped stress-test and refine the model across daily routines, while Aira specialists worked directly with our teams as subject-matter experts to help establish and evaluate our safety guardrails.

Now, you can simply share your camera during a Gemini Live session to hear clear descriptions of your surroundings, ask detailed follow-up questions, and receive proactive, spoken cues to center whatever you want to explore.

Real-time visual assistance designed for how you experience the world

When you share your camera in Gemini Live with Guided Vision turned on, Gemini acts as a conversational partner. If your camera is pointed too high, too close, or slightly off to the side, Gemini provides natural verbal prompts to help you reframe — asking you to pan slowly to the right, tilt downward, or step back so it can get the clear visual context needed to answer your questions.

Everyday moments made simple

Whether locating objects, preparing a meal, or picking an outfit, Guided Vision supports everyday tasks where quick visual verification matters most:

• Reading fine print and complex text: Read small nutrition labels on food packages, decipher appliance dials and washing machine cycles, or read printed menus in dimly lit restaurants.
• Finding and localizing objects: Search your immediate area for everyday items, such as finding a dropped earbud on the floor or locating the black pepper inside a crowded spice cabinet.
• Describing objects and matching details: Ask about colors, patterns, and shapes — such as checking whether a striped shirt pairs with a pair of trousers or identifying the color of a specific jacket.
• Exploring your immediate environment: Get descriptive overviews of unfamiliar spaces, room layouts, or objects placed across a table.
• Conversing naturally in your preferred language: Guided Vision supports conversations across multiple languages. It reflects extensive testing and feedback from blind and low-vision communities across India, Brazil, Singapore, Indonesia, Japan, and more to make descriptions, reframing cues, and answers feel natural and helpful.

Like many accessibility innovations, Guided Vision also provides practical visual assistance for older adults, individuals with low literacy, or anyone trying to read fine print in low lighting.

Quick access across Android

Guided Vision is designed to fit seamlessly into daily routines with minimal friction:

• Gemini App: Open your profile settings in the Gemini mobile app and switch on Use Guided Vision in Live. From there, open Gemini Live and tap to share your camera.
• Accessibility Shortcuts: Under Android Settings > Accessibility > Vision assistance > Guided Vision, configure an Accessibility shortcut to launch Guided Vision using the floating accessibility button, a two-finger swipe gesture, or by pressing both volume keys.
• TalkBack Integration: Users running Google TalkBack can bring up the TalkBack menu with a three-finger tap and select Guided Vision directly.

As with any generative AI technology, Guided Vision is an assistive utility and can make mistakes. It is not a medical device, mobility aid, or white cane replacement, and it’s not intended for navigation, safe-travel guidance, or obstacle detection. Users should continue relying on established mobility aids and safe-travel practices in physical environments.

Guided Vision is now available on Android devices running Android 9 and above in regions and languages where Gemini Live is supported. Make sure your Gemini app and Android system software are updated to the latest versions to get started.

Original source

Gemini 4 Argon delivers frontier performance in complex workflows across real-world software engineering, enterprise knowledge work like legal and finance, and cybersecurity defense.

Today, we’re announcing our new frontier model, Gemini 4 Argon, which is rolling out to a set of trusted cyber defenders through our Fairwind Program. Built to sustain deep reasoning across complex, long-horizon workflows, Argon is fundamentally changing the way we work and build at Google. It delivers frontier performance in complex workflows across real-world software engineering, enterprise knowledge work like legal and finance, and cybersecurity defense.

Safely releasing frontier capabilities at this level requires a phased approach. We are actively engaged in the U.S. government’s voluntary process for pre-release model access while we gradually expand access. We’ll continue to gather feedback from early testers as we iterate on guardrails before making Argon available to developers, enterprises, and consumers as soon as possible.

Argon will launch at an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off input token price.

Gemini 4 Argon is already powering our internal workflows, with thousands of Googlers highlighting the model’s strengths in specialized coding tasks, conducting deeper research, and writing quality. It’s helping teams build faster and push the boundaries of engineering productivity and accelerating breakthroughs:

• Quantum algorithmic optimization: Argon is helping our quantum computing researchers optimize the spacetime resources (qubits × gates) of subroutines that bottleneck important applications. In one example, it beat the published baseline by 40% in a matter of minutes.
• Memory efficiency: A team of Argon agents analyzed fleet-wide profiling telemetry to autonomously identify and apply memory optimizations across Google’s data centers, freeing up over 300 TiB of memory once rolled out, with an estimated 500 TiB to 1 PiB in total savings.
• Large Scale Codebase Migrations and Optimizations: Argon agents are working on migrating C/C++ codebases to Rust across Google — scaling from tens of thousands of lines in core libraries like re2, libgav1 up to 800K+ lines for the Fuchsia Zircon kernel. Given the criticality of many of these systems, such large-scale rewrites are undergoing rigorous automated and manual auditing, emulation testing, and review before rolling out to production.

For libgav1, Google's open source software for decoding video, Argon agents took an existing Rust port and replaced 32K lines of SIMD code by running many rounds of profile-guided experiments, studying the compiler's output, producing safe Rust so the compiler would vectorize it automatically. The end result is a memory-safe video decoder that runs 2.7x faster than the Rust port, with identical video output, bringing it closer to the optimized C++.

To support Gemini 4 Argon’s capabilities across longer, more complex use cases, we are significantly expanding the model’s output token limit to an industry-leading 1M tokens, up from the previous 64K tokens. When the model has the headroom to think deeply and generate hundreds of thousands of tokens in a single trajectory, it adds a new level of depth in reasoning to solve tough problems in one go.

Gemini 4 Argon’s capabilities across coding, reasoning, and multimodality and its ability to sustain long, multi-step tasks enable it to excel across a range of enterprise workflows.

Google engineers have been using Argon for their daily tasks, from everyday debugging to large-scale codebase migrations and algorithm designs. It sets a new state of the art on DeepSWE v1.1 (77.9%), which measures a model’s performance in real-world long-horizon software engineering tasks.

Beyond coding, Argon is the leading model on the Vals Index, which measures economic impact across finance, coding, legal, and tax work, with every sector weighted by its contribution to U.S. GDP. We see similarly leading performance across other domain specific evaluations, like Vals Finance Agent v2 (multi-step financial research) and Harvey’s Legal Agent Benchmark (legal research and drafting). On AutomationBench, Zapier’s benchmark measuring end-to-end execution across core business functions, Argon ranks #1 with a score of 51.3%.

Argon is also uniquely strong when knowledge work requires visual understanding. It’s able to drive professional chart analysis, identify details from long videos, and take action based on a series of documents. For example, on LVBench, which measures long video understanding, Argon is state of the art with a score of 91.7%.

To better equip cyber defenders for the new era of cyberattacks, we trained Gemini 4 Argon to be highly capable at cybersecurity defense. Argon can autonomously find, validate, and patch critical software vulnerabilities. For trusted defenders and our own internal teams at Google, we’ll be releasing Argon without cyber guardrails so they can leverage its full frontier-level cybersecurity defense capabilities.

Wiz is already using Argon for cybersecurity defense through its Scan for Good initiative – a program dedicated to protecting critical public infrastructure for free by finding and remediating high-risk exposures. In an early demonstration of its impact, the model uncovered a critical vulnerability exposing sensitive personal information across healthcare software used by hospitals worldwide, identifying a severe risk that previous frontier models had missed.

On CWE-bench v1, which evaluates the model’s ability to remediate security vulnerabilities, Argon ties for first place with a top score of 68%, building on 3.8 Flash Cyber’s frontier performance on CWE-bench v0.

Gemini 4 Argon demonstrates impressive leaps in vulnerability discovery over 3.8 Flash Cyber. For example:

• On Google’s internal comprehensive vulnerability benchmark, Argon uncovered a wide range of exposures across complex codebases spanning 20 programming languages.
• On Wiz’s internal black-box penetration testing benchmark, which tests a model’s ability to analyze live web systems without source code, Argon outperforms 3.8 Flash Cyber in discovering the attack surface, identifying vulnerabilities, and producing proof-of-concept evidence to validate them.

Before rolling out Gemini 4 Argon broadly, we’re continuing to strengthen critical frontier safeguards across four main areas:

Defending against misuse: To prevent bad actors from using Argon for cyber or chemical, biological, radiological, and nuclear (CBRN) attacks, the model is designed to refuse harmful requests while preserving legitimate, dual-use scientific research, as per our Frontier Safety Framework. We are strengthening the robustness of our safeguards for this launch, including improving our techniques to monitor the model’s internal activations to spot misuse. These safeguards underwent robustness testing by internal and external red teams using a combination of manual and automated attack methods.

Defending against prompt injection attacks: Argon is also our most resilient model yet against indirect prompt injections, where malicious instructions or context are used to hijack a model’s behavior. These are complex attacks that require constant vigilance and multiple layers of defense. Through automated red teaming and adversarial training, Gemini 4 Argon is leading in prompt injection robustness on the Gray Swan’s Indirect Prompt Injection (IPI) benchmark.

Monitoring for misalignment: In order to prevent Argon from stepping out of bounds to try to accomplish a task in a way that goes beyond the user’s intentions, we are deploying misalignment mitigations that monitor Argon’s chain-of-thought and actions and stop execution when necessary.

We used a similar system to monitor our training runs and send alerts to a dedicated incident response team, taking careful precautions against feeding the findings back into training so as to not risk shaping Argon’s reasoning to evade our monitoring. We strongly encourage the rest of the industry to preserve reasoning transparency in these pivotal moments of increased capabilities while navigating alignment risks, so that model thoughts remain helpful in identifying and diagnosing misalignment.

Hardening systems: As frontier models grow increasingly capable, safely testing them requires secure environments that can keep up with the systems themselves. In line with our agent control roadmap, we are hardening our sandboxed environments by isolating and sealing them before high-risk training or evaluations begin. We’re committed to sharing these agent security best practices with our partners to improve security across the ecosystem.

We built Gemini 4 Argon with frontier-level capabilities on coding, knowledge work, cybersecurity defense, and creative writing to be a partner for developers, professionals, and enterprises while they tackle the most difficult problems. We’re grateful for the initial cohort of cyber defenders and trusted testers whose real-world evaluations and feedback will help us strengthen our systems before we release to developers, enterprises, and consumers, starting with paid API customers and Google AI Ultra subscribers.

Original source