AI Models Release Notes
Release notes for leading AI models, APIs and AI platforms
Products (16)
Latest AI Models Updates
- Aug 14, 2026
- Date parsed from source:Aug 14, 2026
- First seen by Releasebot:Aug 15, 2026
August 14, 2026
ChatGPT adds interactive quizzes, lets users change project memory settings, and improves personalized suggestions. It also brings Think on Free and Go to web, smoother dictation and composer behavior on Android, and a public preview of the Linux desktop app.
ChatGPT app experience updates
We’ve made it easier to study with interactive quizzes, manage project memory, and move between typing, dictation, and desktop work.
Learning, Projects, and personalized suggestions
- Practice with interactive quizzes. Ask ChatGPT to quiz you on a topic and answer questions directly in your conversation. Available to all consumer ChatGPT plans and Edu plans on web and mobile.
- Change an existing project’s memory setting. For eligible unshared projects, switch between default and project-only memory without starting a new project. Shared projects remain project-only.
- Get suggestions that fit how you use ChatGPT. Eligible paid users can see homepage suggestions based on their conversation history and connected tools.
Web
- Use Think on Free and Go. Select Think on the web when you want ChatGPT to reason through a harder question.
Android
- Switch between typing and dictation. The dictation button stays available after you add text or attachments, so you can continue building the same message.
- See suggestions that stay in sync. Composer suggestions update as you type or delete text.
- Keep typing with a physical keyboard. Connected keyboards keep the composer focused, preventing interruptions while you write.
Linux desktop
- Use ChatGPT and Codex on Linux. The desktop app is now available globally in public preview on Ubuntu 24.04 LTS and 26.04 LTS, Debian 13, and Fedora 43 and 44.
- Work in the built-in browser or Chrome. Browser actions are available in the Linux app; controlling other desktop apps is not yet supported.
Download ChatGPT for Linux.
Edit memory settings for existing projects
You can now change a project’s memory setting after you create it. Open the project, select the three-dot menu, choose Project settings, and select Default memory or Project-only memory under Memory.
With project-only memory, ChatGPT can use conversations from the same project for context, but it won’t reference memories or conversations from outside the project. Information from the project is also kept out of memory used in chats outside the project. Changes may take a few hours to take effect.
This update is available on all ChatGPT plans. Shared projects continue to use project-only memory and cannot be switched to default memory. ChatGPT Work is not available in projects using project-only memory.
Learn more about Projects in ChatGPT.
Original source - Aug 14, 2026
- Date parsed from source:Aug 14, 2026
- First seen by Releasebot:Aug 15, 2026
How Claude’s text watermark works
Anthropic adds watermarking to future Claude text outputs to support EU AI Act compliance, with no visible impact on quality, speed, or pricing. The company also says a watermark detection API is coming soon and that older Claude models will gain support over the coming months.
Future Claude models will generate text that contains a watermark. This is a way of determining the likelihood that Claude was involved in writing the text, and we, along with several other major AI providers, are implementing this change to comply with the EU AI Act.
In this article, we share answers to some of the questions we’ve received about how our chosen watermarking method works, whether it affects Claude’s outputs, and why we’re making this change. To summarize:
- We use a method of watermarking that does not have any practical impact on the quality or content of Claude’s outputs;
- The difference between watermarked and un-watermarked text will not be distinguishable to readers;
- Nothing is added to the text and there are no hidden characters;
- Watermarking doesn’t require extra tokens, and will not be more expensive;
- Watermarking carries no identifying information and can’t be traced to a specific person, organization, or chat;
- Watermarking won’t be specific to Claude. As of August 2, the EU requires AI providers serving its market to mark AI-generated content. Other major model developers have signed the same Code of Practice and will be implementing their own watermarks.
What is watermarking?
Large language models like Claude work by generating one word at a time. Each time the model decides on the next word, it chooses among a list of potential candidates, ultimately selecting the most sensible or likely based on the preceding text. Take the sentence “The weather today was cold and…”. The next word is very unlikely to be “sugary.” But it is quite likely to be “overcast” or “grey.” Under most circumstances, it doesn’t matter much to the reader which of these latter two words the model ultimately chooses—the meaning of the sentence is largely the same either way. In cases like this, the choice is settled by a random number.
Watermarking uses low-stakes choices like these—which occur many times over a piece of generated text—to leave a pattern in Claude’s responses. That pattern is undetectable to the reader, but is detectable to anyone who has a key that encodes it. When watermarking is used, choices are still made at random, but the source of the randomness is different. Instead of using an arbitrary random number generator to pick the next word, watermaking uses the key and a few words that come before to settle what word the model should pick. That is, the words that Claude picks are still random, but now, one can check the sequence of words and see if it’s consistent with the choices Claude would make if it was using the key. If it is, one can assign a probability that the text was generated by Claude.
Importantly, it isn’t that the model will now always be biased toward overcast or grey. Just as with non-watermarked text, overcast might be selected in one sentence, grey in the next, depending on the words that came before. And it’s not the case that the watermarking method pushes Claude to choose a word it wouldn’t have considered anyway (for instance, it wouldn’t make Claude pick a word like “nubilous”—an obscure1 synonym for overcast or grey that Claude almost certainly wouldn’t use under normal circumstances).
How does watermarking affect Claude’s outputs?
Watermarking does not impact the quality of Claude’s output. To a reader, a watermarked response is indistinguishable from an unwatermarked one (in this way, AI watermarks differ substantially from their namesakes on banknotes, other physical objects, and some digital documents, which are visible to the naked eye).
In internal testing, we’ve seen no impact of watermarking on the content, level of creativity, or readability of Claude’s text. In the SynthID-Text paper, which introduced the technique we use, Google DeepMind tested this impact by serving a model that used watermarking to a portion of their Gemini traffic and comparing thumbs-up and thumbs-down ratings. They found no statistically significant differences from the unwatermarked model. And in a controlled study, human raters comparing watermarked and unwatermarked answers side-by-side saw no difference in quality.
A useful analogy is to imagine you’re playing a game like Monopoly. On each turn, each player moves a random number of spaces around the board according to the roll of a die. Suppose that, instead of rolling the die to get this randomness, we decided to use a book of the digits of pi.2 We start from a randomly-chosen digit (say, the 1,012,845th after the decimal place, which happens to be a 6), and from that point on each player simply uses the next digit in the sequence as their next “roll”.
For all intents and purposes, the moves are still random: it makes no difference to the players—or to the outcome of the game—whether the randomness comes from pi or from dice rolls each time. But if we could see the sequence of all the moves after the game (and we knew the value of pi), we could work out whether this was a game that likely used pi to determine its moves. The game that used pi is, in a sense, “watermarked”.
It’s the same for Claude-generated text. Watermarking doesn’t change the meaning or experience for the person reading it, but if you wanted to check after the fact whether the text was likely generated by Claude, the watermark allows you to do so.
Which specific method of watermarking do you use?
Claude’s text watermark is a version of the SynthID-Text approach published by Google DeepMind in a Nature paper in 2024. It belongs to a family of approaches that go back to a proposal by Scott Aaronson in 2022, all of which share the same design principle that we described above—the watermark only changes the source of the randomness used to pick among words.
There are limitations to the effectiveness of watermarking. Using our key, one can only answer the question “What is the likelihood this was partly written by Claude?” It doesn’t confirm whether the text was human-written, and it can’t tell whether the text was written by a different AI (even if that other AI uses watermarking, it would have a different key; it might also use a different watermarking method altogether). Detecting a watermarking also doesn’t work well on small samples, where there are fewer word choices and thus less information to go on. As a passage increases in length, confidence about Claude’s involvement increases too.
Watermarking is sparser on factual passages where there are fewer choices that can be made without decreasing the accuracy of the text. For example, take the sentence “Isaac Newton’s most famous work was called Principia…”. It really matters whether the next word is “Mathematica” (it’s the only right answer), so the watermark would have nothing to act on. The same is true for proofreading. If you hand Claude a piece of writing and ask it to edit only the grammar and punctuation and nothing else, the watermark can only live in the handful of corrections, which might be too few to register.
What about cases where Claude has proofread or edited human text?
The watermark only applies to words Claude chooses. When Claude proofreads text written by a person, what it gives back has generally only been lightly edited; because nearly all the words are the person’s, there’s very little (if anything) for the watermark to attach to. Depending on the length of the text and how heavily Claude has edited it, those changes might not be enough to make Claude’s involvement detectable. The more Claude writes, the more decisions it has to make, and the more space there is for a watermark.
What about code?
As we noted above, AI watermarking takes advantage of decisions where either choice of a word would be equally good. Where an exact output is required—where there isn’t a choice, and something would be factually wrong or a piece of code would break if a different term was chosen—the watermark isn’t applied.
For example, once the model has written “2 + 2 =”, there is a very clear best choice for the next token (if the model is completing the sum, there isn’t an answer that’s equally as good as “4”; if it’s talking about George Orwell’s Nineteen Eighty-Four, there isn’t an answer that’s equally as good as “5”). The “nudge” of the watermark wouldn’t be applied here. For the same reason, code—which in very many cases has to be exact—has generally less watermarking than some other forms of text.
Having said that, in areas where there is an arbitrary choice between particular words or terms within the code, the watermark can be used, such as comments within code. But by definition, it will have a negligible effect on the actual code produced.
What does this mean for users?
Does this slow the model down, or make it more expensive?
No. Watermarking has a negligible impact on the speed of models, and because it produces no extra tokens, the model is the same price to serve and use.
Can a watermark be traced back to me or my organization?
No. The watermarking applies to Claude and its outputs. It doesn’t identify anything to do with individual users. There’s nothing in the watermark, or its key, that would allow anyone to recover any information about the user, their organization, or their chats with Claude.
Why are you watermarking Claude’s outputs?
We’re implementing watermarking to comply with the EU AI Act. Anthropic, along with several other major AI model providers and around 190 total signatories, signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026. This requires AI system providers to use methods of “marking” AI-generated text. We’re applying watermarking globally at launch because we don't yet have a durable way to scope it by region. However, we will continue to evaluate different approaches, and will share updates when we have them.
Other questions
How do I check if a piece of text was written by Claude?
We will soon be offering a watermark detection API. We’re in the process of working out the details of its implementation.
What about images and other files?
When Claude produces a file of a supported type (such as a .png, .jpg, or .svg), it will attach a content credential in the form of a small, cryptographically signed note in the file’s metadata, saying that the file was made or processed with Claude. This is an open industry standard called C2PA—the same used by camera manufacturers and in photo-editing software to record where an image came from. Any C2PA-aware tool can read it; we’ll be providing our own where you can drop a file and check.
This metadata label is very different from a watermark. Nothing in the file changes—it is not embedded or hidden. As with text, the credential only says Claude was involved in producing the file; it doesn’t include any identifying information.
Can’t someone just edit the text to get around the watermarking?
To some extent, yes. Light editing probably won’t remove the watermark completely; a complete rewrite where every word is replaced will. In the latter case, of course, it’s arguable whether the text can any longer be described as AI-generated.
What does a watermark actually prove?
A watermark can only determine that Claude was likely involved with the content at some point. It cannot distinguish “Claude wrote this” from “Claude heavily edited this.”
Do watermarks apply to translations?
Yes. A translation produced by Claude carries a watermark, because in this case every word is chosen by Claude.
What about older Claude models?
The EU law includes a transition period for Anthropic models launched before August 2, 2026, and we’re working to add watermarking for those models as well. This will be rolled out over the coming months.
How does this differ from AI detection software, like Pangram?
AI detection software uses a different method, because the companies that provide it don’t have our key. Among other things, those services look at aspects of the text like the subtle (and not-so-subtle) “tells” that often appear in AI’s phrasing. For example, AI models appear to be fond of the construction “this isn’t [X], it’s [Y]”, and use the word “quietly” a lot more than you might expect. Picking up on these patterns is fundamentally different from checking for a watermark.
Does this change who owns a given output, or who is legally responsible for it?
No. A watermark only helps test whether Claude might have produced or processed the content. It doesn’t say anything about ownership or authorship, and doesn’t change a user’s rights under our terms. We only apply the watermark when Claude was involved in processing the content or file.
Original source All of your release notes in one feed
Join Releasebot and get updates from OpenAI and hundreds of other software products.
- Aug 14, 2026
- Date parsed from source:Aug 14, 2026
- First seen by Releasebot:Aug 14, 2026
Grok 4.6 in GitHub Copilot
xAI adds Grok 4.6 to GitHub Copilot, bringing its latest coding model to VS Code and GitHub workflows. Developers can select it in the model picker, with enterprise enablement available in Copilot settings and access also offered through the SpaceXAI console.
Grok 4.6, our latest coding model, is now available in GitHub Copilot.
Grok 4.6 is now live in GitHub Copilot for the millions of developers who work in VS Code and across GitHub every day.
Getting started
GitHub Copilot covers several products, including cloud agents, the Copilot CLI, and the VS Code IDE. Open the model picker and select "Grok 4.6" to start using our latest coding model. Some businesses and enterprises will need to enable the model from Copilot settings.
Grok 4.6 is also available from the SpaceXAI console at $2 per million input tokens and $6 per million output tokens. For more on the model, including coding performance, read the Grok 4.6 announcement.
Original source - Aug 13, 2026
- Date parsed from source:Aug 13, 2026
- First seen by Releasebot:Aug 14, 2026
The builder’s guide to GPT‑5.6
OpenAI releases GPT-5.6 with stronger agent performance at much lower cost, plus new Responses API controls for reasoning continuity, native multi-agent orchestration, programmatic tool calling, and better prompt caching to help teams build faster, more efficient agents.
Technical lessons from startups in production
GPT‑5.6 sets a new standard for price-performance
The GPT‑5.6 model family makes frontier-level agent performance dramatically more affordable, while also advancing the frontier of what is possible.
In this guide, we show how startups are using smarter model selection and new API controls that help with reasoning continuity, multi-agent orchestration, and programmatic tool calling to build faster, more capable agents at a fraction of the cost.
A better out-of-the-box experience
Since GPT‑5, each model generation has sought to tackle longer-horizon tasks with fewer tokens. GPT‑5.6 continues that trajectory: stronger agent performance, lower costs, with minimal changes to the underlying harness.
The improvements in top-line cost efficiency are compounded with increased accuracy at lower reasoning efforts. For example, on Agents’ Last Exam, GPT‑5.6 Sol at “low” reasoning outperformed GPT‑5.5 at “high” reasoning when the harness was kept constant. We’ve seen similar success stories in production testing where startups report seeing significant cost improvements across a range of workflows by reducing the reasoning effort from the prior defaults.
“We dropped GPT‑5.6 into our harness, and low reasoning effort gave us our best results. It knew when the data just wasn’t there, didn’t chase bad leads, and got to the right answer with fewer tokens.”
— Izzy Miller, AI Research Lead, HexModel Selection
Historically, upgrading to a flagship model at the highest reasoning available has been the best option for long-horizon use cases. This has been in large part due to these models being significantly more capable than cost-optimized models at handling longer contexts and tool calling. This has changed with the 5.6-family: with more test-time compute, Luna and Terra can often perform similar to GPT‑5.4 and 5.5 while being significantly cheaper.
Consider tasks in BrowseComp: a search-based benchmark that tests a model’s ability to search for obscure facts. Three months ago, GPT‑5.5 (Extra High) scored 84.36% on this benchmark for a total cost of $33.27. At launch, GPT‑5.6 Luna (Extra High) delivers essentially the same performance, scoring 84.04% at a cost of $1.33. We’ve since reduced prices further.
The smaller 5.6-family models are a strong fit for high-volume workloads, latency-sensitive interactions, and repeated steps within agentic workflows. For example, if you’re operating a legal-tech startup that parses handwritten memos prior to agentic analysis, instead of using a frontier model for the entire use case, you can now use Terra or Luna for extraction and register significant cost savings.
Evolving the Responses API to architect more efficient agents
In addition to making GPT‑5.6 more performant out of the box, we also shipped new primitives to the Responses API to unlock further gains. We trained GPT‑5.6 end-to-end with three complementary architectural interventions that enable agents to operate more efficiently:
- Reuse work already performed: by allowing reasoning to be persisted across model turns and using native compaction to compress long-running conversations, the model can maintain coherence in its work across longer task horizons without getting confused or having to reconstruct prior context.
- Parallel decomposition where appropriate: using native multi-agent orchestration allows coordinating multiple agents across parallel workstreams to finish complex tasks faster.
- Move deterministic work into code: using programmatic tool calling to filter, aggregate, and orchestrate tool outputs outside the model’s context window, reserving model tokens for judgment and reducing cost, latency, and context rot.
Used together, the difference can be dramatic. For example, on ARC-AGI-3, GPT‑5.6 Sol scored 13.3% with the standard harness. After enabling retained reasoning and compaction, however, the score jumped to 38.3%—while using roughly 6× fewer output tokens. No changes to the model, but nearly three times the performance. You can read more on our ARC-AGI-3 harness investigation here.
Programmatic Tool Calling
Agentic workflows often involve two kinds of work:
- Tasks that require judgment
- Work that mostly requires moving, filtering, and combining data
When an agent retrieves 100 filings, filters them by date, and identifies relevant transactions, the model shouldn’t have to reason over every intermediate result in its context window. Programmatic Tool Calling lets GPT‑5.6 write JavaScript to orchestrate tools, run independent calls in parallel, and process their outputs outside the context window. The model is left to focus on what requires intelligence: applying judgment.
“For financial research, the hard part is reliably pulling filings, coordinating tools, and working through the numbers. In our evaluations, GPT‑5.6 using Programmatic Tool Calling matched our rubric quality while using 21% fewer input tokens. That’s the difference between an agent that can discuss financial research and one that can actually carry it out.”
— Alex Wang, Applied AI, RogoMulti-agent
On complex, parallelizable tasks, distributing actions and reasoning across multiple agent workstreams enables faster task completion as well as higher intelligence. In these setups, the primary agent is responsible for orchestrating the subagents and delegating tasks to them. The subagents pursue their objectives in parallel and finally pass back their output to the primary agent for final synthesis. Teams can start leveraging multi-agent natively by enabling multi-agent in the Responses API. This is also how the ultra capability setting in ChatGPT works.
“Qualia runs teams of agents on open-ended research problems, and GPT‑5.6 Sol just clicked. It showed a marked improvement over GPT‑5.5, finished faster than almost every other model we tested, and quickly became our go-to OpenAI model.”
— E Chi, Founder, Quadrillion“GPT‑5.6 is the best orchestrator we’ve seen from OpenAI. We threw six specs at it at once (writing, building, and talking through all of them) and it kept track of everything without the quality falling apart.”
— Jon Bell, Co-founder and CPO, ObviousAlthough GPT‑5.6 has a strong sense of the appropriate number of subagents and when to spawn them, multi-agent behavior is very steerable. Instructing the model on when to invoke subagents can increase the likelihood of spawning agents only in situations where the additional token expenditures would result in better performance.
Prompt Caching
Across the entire family of models, the prompt cache TTL has been extended to a minimum of 30 minutes and cache breakpoints can now be set deterministically within a model’s context window. This has enabled startups to significantly improve their cache hit rate.
“We added cache breakpoints and workspace-specific keys to a shared 29,000-token prompt and cut uncached input by 28%. The 30-minute cache window was a big unlock too: our agents could reuse the same context across runs instead of starting from scratch.”
— Lorenzo Gentile, AI Engineer, PloyIn addition to setting cache breakpoints, continuing to use an appropriate prompt_cache_key increases the likelihood of requests landing on the same inference engine as one that previously served the same prefix, thereby reducing latency.
Conclusion
What stands out across these examples is how much the economics of building agents have changed.
Use cases that once required a frontier model at every step can now achieve comparable or better results at a fraction of the cost by using smaller models, tuning reasoning effort, and making efficient architectural choices.
We're excited to see what you all build!
Original source - Aug 13, 2026
- Date parsed from source:Aug 13, 2026
- First seen by Releasebot:Aug 14, 2026
Claude Tag now reads even more of the room
Claude adds smarter Slack collaboration with Claude Tag, using full channel context, memory, and standing instructions to decide when to jump in or stay quiet. It now responds more accurately and faster, with proactive replies available today at no extra cost.
Claude Tag lets you add Claude to a Slack channel, where it works alongside your team. Claude responds when you @-mention it, or proactively when it thinks it can be helpful.
Before, Claude only saw one message at a time, so it made decisions to act proactively based on what was in front of it, but not the wider context of what was around it.
Now, Claude uses context from across the channel, as well as its memory and the standing instructions you have given it, to determine when to contribute to the conversation.
As a result, Claude is now roughly 30% better at determining when, and when not to, proactively respond.
This update comes at no additional cost today. While holding more context does increase Claude Tag’s usage, the additional context Claude Tag holds does not count toward usage or spend limits on any plan.
From passive responder to active participant
Previously a lightweight classifier decided when Claude should act. It looked at each new message on its own and made one yes-or-no call.
For example, here are two engineers chasing the same bug from opposite ends. Neither has a free hour to run it down, and neither message asks for anything.
Priya has a theory. Devon has the evidence. Neither message is for Claude, and neither asks for anything.
Read one at a time, neither message is for Claude, so the classifier correctly does nothing, twice. Read together, there's an obvious piece of work sitting there. One engineer has a theory, the other has the evidence for it, and nobody has time to check.
With the classifier removed, Claude uses context across the channel to make one of four moves:
- Reply inline, when the answer is short, verifiable, and something the channel doesn't already know.
- Start deeper work in a thread, when a message deserves real time.
- Route the message to work it has in flight, when it adds to a workstream Claude already has open.
- Say nothing, when nothing is called for.
Here's the same conversation with Claude Tag using additional context. Claude picks the second move, even without being @-mentioned. It sees Priya's hypothesis and Devon's evidence, opens a thread with the investigation already running, and pulls both engineers in. It acts within the boundaries of the permissions, tools, and scope you have configured.
Same thread, two minutes later. Claude reads the two messages together and starts the work. No @-mention.
The conversations aren't walled off from each other. So when Devon posts an update, it lands in the right workstream. When two investigations turn out to be the same bug, that connection gets made.
Claude now looks at all messages to understand the full context of the channel, to more accurately determine if it should participate in a conversation unprompted.
How Claude decides when not to speak
An annoying agent is worse than an unhelpful one. We built Claude Tag to speak up only when it's useful, and in most channels, on most messages, that means saying nothing.
We do this by grading Claude’s channel-by-channel choices against a rubric based on principles like how useful the comment is, how confident Claude is in the response, and whether there is a person better suited to respond.
Claude also knows when to stop paying attention, similar to how people navigate Slack. It follows a few channels closely while paying less attention to others until someone tags it in. In a channel where, message after message, Claude keeps concluding it has nothing to add, it goes to sleep. A @-mention wakes it instantly.
You can also steer its response behavior in plain language: "Never respond here unless someone tags you," or "Feel free to jump in on anything about the deploy pipeline."
And if you'd rather Claude only spoke in a channel when someone tags it, any member can switch ‘Respond automatically’ off.
The first reply is faster
The additional context also allows Claude to respond more quickly. It acknowledges you in seconds instead of operating silently while it starts up. The work itself takes as long as it always did; what's gone is the silent first minute when you couldn't tell whether it heard you.
Live today
This update is now available across Claude Tag, available for Claude Teams and Enterprise customers. You can get started here. Claude now acts as a more effective collaborator, one that can follow the conversation, decide for itself when to act, and when to stay out of the way.
Add Claude to one channel and watch what it adds to your conversations. Learn more about Claude Tag.
Original source - Aug 13, 2026
- Date parsed from source:Aug 13, 2026
- First seen by Releasebot:Aug 14, 2026
- Modified by Releasebot:Aug 14, 2026
August 13, 2026
ChatGPT adds Google Drive in Library and a new macOS Computer History feature. Users can browse Drive files, pull them into chats, work with Docs, Sheets and Slides side by side, and ask ChatGPT to reference past app and web activity on Mac while keeping the data under user control.
Google Drive is now in Library
If you have the Google Drive plugin connected, you can now see and browse your Google Drive files and folders directly from Library, including items shared directly with you. You can also quickly pull up a Drive file from the composer or with @mentions and add it to any chat—without uploading it again.
When you’re working with a file, you can now keep Google Docs, Sheets, and Slides open beside the conversation while asking ChatGPT to summarize, analyze, compare, or create something new from them. You can also select a folder and ask ChatGPT to work across the files it contains. Your content stays connected to Drive, so it’s easy to return to the original. Where supported and authorized, ChatGPT can update the source file directly.
Availability
The initial experience includes My Drive and files and folders shared directly with you; Shared Drives aren’t included yet. Some Google Drive editing and collaboration features aren’t yet available in ChatGPT.
Rolling out to Plus, Pro, Enterprise, Edu, Healthcare and Business users on the web in both the Chat and Work toggles. Mobile support will follow.
Computer History for macOS
Computer History is an optional feature in the ChatGPT macOS app that lets ChatGPT and Codex reference selected activity from apps and websites, helping users continue work without re-explaining every detail. It records interaction events—not screenshots, screen recordings, microphone input, or system audio—and private browsing is not included.
Computer History is off by default and available to Pro, Business, and Enterprise users. Business and Enterprise admins must grant access before members can opt in. Users can pause Computer History, choose included apps and sites, and inspect or delete timeline items. Note that Computer History is not currently available in the EEA, UK or Switzerland.
Original source - Aug 13, 2026
- Date parsed from source:Aug 13, 2026
- First seen by Releasebot:Aug 13, 2026
Introducing Gemini 3.7 Flash
Gemini introduces Gemini 3.7 Flash, a faster, more capable workhorse model for coding, agents, web development, and knowledge work, with stronger reasoning, better developer experience, updated safety safeguards, and an introductory lower price. Gemini Spark also moves to the new model.
Better intelligence for complex workflows
Today, we’re building on the progress of our widely used Flash series by introducing Gemini 3.7 Flash, our most intelligent workhorse model yet for coding and agents.
This release comes just three weeks after Gemini 3.6 Flash, and is a direct result of developer feedback and algorithmic innovations that we look forward to bringing to future models. 3.7 Flash delivers substantial improvements across software engineering, knowledge work, and web development workflows — with an introductory price of half the original 3.6 Flash cost per million tokens.
3.7 Flash shows strong gains over 3.6 Flash in coding tasks like debugging and issue resolution. It also achieves higher first-pass code accuracy and has improved performance in generating production-ready code as seen in FrontierCode 1.1 Main (43.6% vs 34.4%) and DeepSWE v1.1 (65.3% vs 49.0%).
In web development, 3.7 Flash generates more functional layouts and feature-complete apps in fewer prompts. For UI generation, the model shows high design adherence and parity based on a reference input, whether it’s a screenshot, an image, or a full design system. It outperforms 3.6 Flash on Arena.ai’s WebDev Arena with an Elo score of 1588 vs 1538.
For knowledge-dense fields like finance, law, and biosciences, 3.7 Flash delivers improved reasoning and accuracy. It significantly outperforms 3.6 Flash on the GDP.pdf benchmark (34.0% vs 22.0%), an eval for testing a model’s ability to process complex documents. It also surpasses 3.6 Flash in AutomationBench, demonstrating it can more effectively complete real-world business workflows (30.4% vs 17.0%).
Better developer experience and price
Gemini 3.7 Flash delivers a noticeably improved developer experience over 3.6 Flash. It better adapts to roadblocks, clarifies intent when needed, and follows instructions with greater fidelity. It thinks more diligently, putting in more effort into multi-step planning and tool calls. A more disciplined execution means less manual oversight and fewer retries across engineering workflows.
3.7 Flash is available through the end of the year at an introductory price of $0.75/1M input tokens and $3.75/1M output tokens. This price combined with the enhanced model performance enables developers and customers to scale production-ready agents cost effectively.
Early customer feedback is highlighting 3.7 Flash’s performance and precision, achieving results that are significantly better than 3.6 Flash at a low cost.
Improving Gemini Spark with 3.7 Flash
Gemini Spark, available to Google AI Pro and Ultra subscribers in over 160 countries, will be using Gemini 3.7 Flash starting today. We launched Spark at I/O as your personal AI agent that runs 24/7, taking action on your behalf while under your direction. This model update makes Spark more efficient for knowledge work with improved tool use for Google Workspace apps, delivering improved accuracy and output quality for complex, multi-skill workflows.
Built with safety in mind
We continually work to improve the coverage and robustness of Frontier Safety safeguards. Gemini 3.7 Flash is shipping with updated safeguards against misuse in the domains of Chemical, Biological, Radiological, and Nuclear (CBRN) and cyber offense, while enabling beneficial use cases, in accordance with our approach to bioresilience and our cyber program.
For more information, see the 3.7 Flash model card.
Try it today
- Developers: Explore agent-first workflows in Google Antigravity or start building today in the Gemini API via Google AI Studio and Android Studio. Get started with our developer guide.
- Enterprises: Access 3.7 Flash in Gemini Enterprise Agent Platform and the Gemini Enterprise app.
- Individuals: Available via Spark, your 24/7 personal agent in the Gemini app for Google AI Pro and Ultra subscribers in supported countries.
- Aug 13, 2026
- Date parsed from source:Aug 13, 2026
- First seen by Releasebot:Aug 13, 2026
Gemini Omni Flash
Gemini introduces Gemini Omni and its first release, Gemini Omni Flash, bringing conversation-style video generation and editing from any input for creators.
We recently introduced Gemini Omni, a new model that lets you create anything from any input. The model’s first release, Gemini Omni Flash, makes generating and editing videos as easy as having a conversation.
To get the inside scoop, we sat down with three Google DeepMind team members who work on Omni: research scientist Mohammad Babaeizadeh, product manager Anish Nangia, and research engineer Sarah Xu. They shared their vision for Omni, explained why they started with video generation, and broke down what exactly makes Omni so helpful for creators.
They also let their imaginations run wild. Could Omni swap their hairstyles? Turn them into sloths? Materialize Sarah’s cat on the table in front of them? “There is no ceiling to this,” Mohammad said.
And what’s next for Omni? “It’s only going to get better from here,” Sarah said.
Watch the full interview below to learn more about Omni.
Original source - Aug 13, 2026
- Date parsed from source:Aug 13, 2026
- First seen by Releasebot:Aug 13, 2026
Previewing Ultrafast mode: GPT‑5.6 Sol at up to 14X the speed
OpenAI introduces Ultrafast, a new API service tier for GPT-5.6 Sol that runs up to 14× faster than Standard processing. Powered by Cerebras, it brings real-time speed to coding, support, research, commerce, and other time-sensitive workflows in limited preview.
A new speed class for frontier intelligence, turning speed into a competitive advantage.
Today, we’re sharing an early look at Ultrafast, a new service tier that runs GPT‑5.6 Sol up to 14× faster than Standard processing, launching first in the OpenAI API. Powered by Cerebras, Ultrafast generates up to 750 output tokens per second, bringing our most intelligent model to products and workflows where every second matters.
With GPT‑5.6, we’re pushing the frontier on what our models can do and making them more efficient across every layer of our stack. Those improvements have made advanced intelligence more affordable and more useful to more people. Until now, getting real-time speed typically meant choosing a smaller or more specialized model. Ultrafast points to progress in a new direction: more useful work per second.
When speed no longer requires giving up intelligence, AI can move into the most time-sensitive parts of a business and new kinds of work become possible. We’ve already seen some encouraging scenarios for Ultrafast:
- Incident response and reliability: When a critical system fails, analyze application logs, recent code changes, and engineer reports to identify the likely cause and help prepare a fix while the outage is still unfolding.
- Financial research and security: Analyze market signals, assess transactions, and identify suspicious activity while conditions are still changing.
- Customer support and voice: Resolve complex customer issues in real time without interrupting the conversation, even when finding the answer requires multiple steps or systems.
- Commerce: Answer product questions, check inventory, personalize recommendations, and resolve checkout issues while the shopper is still deciding, before hesitation becomes an abandoned cart.
- Live research and experimentation: Turn research that previously took an overnight run into an interactive working session, letting teams test an idea, examine the results, adjust their approach, and run another experiment without breaking their flow.
During the preview period, we’re working with an initial group of customers to understand where this speed makes the biggest difference, and how those learnings can inform our products over time. If your business requires frontier intelligence at the highest speed, you can sign up to get notified when access expands.
GPT‑5.6 Sol Ultrafast and standard build a working 3D warehouse simulator from the same text prompt, side by side.
What early customers are experiencing
We’ve been testing GPT‑5.6 Sol on Ultrafast mode with an initial group of companies across coding, commerce, financial research, support, and other interactive applications. Starting with business workflows lets us study these conditions in real production environments. Their early work is helping us understand where an order-of-magnitude change in speed creates the most value and how products change when the model can keep pace with the person using it. We will use these findings to guide deployment as capacity grows.
“ The increase in speed brought by Cerebras is impressive. It enables different ways of using the models, and makes it practical for developers to work in a more focused and productive way alongside them. ”
—John Crepezzi, AI Assistants, Jane StreetHow OpenAI is using Ultrafast
Inside OpenAI, a group of developers has been testing GPT‑5.6 Sol on Ultrafast mode to understand which workflows benefit from frontier intelligence that can answer in real-time.
Incident response is one example where our team is using Ultrafast. When an alert fires, engineers need to build an accurate picture while the system and the evidence are still changing. Teams use it to quickly read logs, analyze traces, synthesize conversations, identify the next checks, and help prepare or validate a fix—all in a fraction of the time with the intelligence of Sol. It reduces the delay between observing a signal, testing a hypothesis, and choosing the next action, while engineers remain responsible for judgment and deployment.
For research, our team uses Ultrafast to rapidly search knowledge sources, query data, and quickly gather, organize, and summarize information across connected tools. A common workflow in research is for our team members to launch a batch of experiments over night, and review the results in the morning. With Ultrafast, we see this loop tightening to support multiple iterations during the workday instead.
Powered by Cerebras
Ultrafast marks the next step in our partnership with Cerebras to bring ultra-low-latency inference to OpenAI’s platform. Now, with GPT‑5.6 Sol on Ultrafast mode, Cerebras is supporting OpenAI’s most intelligent model, delivering up to 750 output tokens per second, enabling businesses to build more responsive products, make faster decisions, and bring powerful AI directly into their most demanding workflows.
Availability
GPT‑5.6 Sol on Ultrafast mode is available in a limited preview today to a select group of customers. We’ll expand access as capacity grows. Sign up for updates.
Original source - Aug 13, 2026
- Date parsed from source:Aug 13, 2026
- First seen by Releasebot:Aug 13, 2026
MiniMax Music 3.0: Next-Generation Open-Weights, Production-Ready & Versatile Music Model
MiniMax introduces Music 3.0, a next-generation music generation model that turns creative concepts and optional lyrics into complete songs up to five minutes long with richer arrangements, clearer sound, and more natural vocals.
Music 3.0: Technical Architecture
Today, we are introducing MiniMax Music 3.0, our next-generation music generation model. Given a creative concept and optional lyrics, the model composes, arranges, performs, and produces a complete song in a single generation.
Music 3.0 focuses on the aspects of music creation that are hardest to capture with a simple prompt: understanding the creator’s expressive intent; sustaining that intent across a complete song of up to five minutes; rendering instruments with clarity and physical realism; and generating vocals that sound performed rather than synthesized.
To achieve this, we redesigned the entire generation pipeline—from music description and language modeling to audio rendering. Fine-grained temporal descriptions capture the evolution of emotion, instrumentation, and vocal delivery. A global–local Hybrid-LM jointly models long-range structure and acoustic detail. Multi-layer residual vector quantization (RVQ), hidden-state fusion, flow matching, and a Flow-VAE further improve output fidelity. Together, these advances deliver three major upgrades: more accurate interpretation of creative intent, more complete and varied arrangements, and clearer, more natural sound.
Music 3.0 comprises three interconnected core components: the tokenizer, the Hybrid-LM, and the synthesis stack. They address, respectively, how musical information is represented, how long-range structure and local detail are modeled, and how audio is reconstructed at high fidelity. The complete workflow is shown below.
Multi-Layer RVQ: Separating Core Structure from Acoustic Detail
An eight-layer RVQ represents musical information hierarchically. The first layer captures core semantics and structure, while the remaining seven progressively encode acoustic residuals. In staged training, the first layer first learns a stable information backbone; all codebooks are then trained jointly to model fine-grained sound. This design provides a more stable discrete representation for long-sequence prediction and avoids overloading a single token layer with both structural and fidelity-related information.
Hybrid-LM: Joint Modeling of Global Structure and Local Acoustics
The 8B Global LLM is initialized from Qwen3.5-8B and performs frame-by-frame prediction of semantic tokens while modeling global context. A randomly initialized 0.6B Local LLM predicts within-frame acoustic tokens along the depth axis. Training proceeds in two stages: global alignment first establishes the Global LLM’s ability to predict musical semantics, followed by full-parameter joint training of the global and local models. This hierarchical collaboration enables the model to preserve song-level structural stability while resolving acoustic detail within each frame.
Hidden-State Fusion: From Discrete Prediction to Continuous Audio Rendering
Conventional systems typically feed discrete acoustic tokens directly into a decoder. Music 3.0 instead fuses continuous hidden states from the Global LLM and Local LLM and uses them to condition a 2.4B flow-matching module. A 123M Flow-VAE then decodes the audio. The complete path is:
Fused LLM features → Flow matching → VAE hidden states → Flow-VAE decoder → Final audio
This design directly connects the language model’s structural understanding with the acoustic model’s audio reconstruction through continuous representations. It preserves long-range consistency while improving pronunciation accuracy, instrumental coherence, and fine-detail fidelity.
Understanding Creative Intent
AI-generated music can sound like a complete song and still drift away from the original brief. Specified instruments may gradually disappear from the arrangement, the intended emotional character may weaken as the song develops, and a requested vocal style may appear in only one section instead of remaining coherent throughout the piece.
Music 3.0 introduces a more expressive music-description framework. Rather than summarizing an entire piece with a single global label, it uses Structured Captions to describe music at fine temporal granularity. These captions specify genre, tempo, time signature, key, use case, and production character, while also tracking emotional contour; the entry and exit of primary and supporting instruments; the development of groove and low-end energy; and section-level changes in vocal delivery, harmony, and vocal effects.
These descriptions convert subjective listening impressions into a professional arrangement framework that the model can learn and execute. As a result, the model can translate creative language into concrete musical expression more accurately, maintain a coherent musical identity as the song unfolds, and still introduce the necessary dynamic variation.
To make professional music creation more accessible, we also developed a template-based Prompt Enhancement System. It selects appropriate language from a curated library of Structured Caption templates and applies established musical terminology and arrangement principles to expand a simple user description into a detailed, musically coherent instruction. Creators can therefore exercise precise control without mastering specialist vocabulary—whether they want an intimate, restrained unplugged performance; a late-night R&B track driven by rolling hi-hats and deep 808 bass; or a cinematic instrumental that grows from introspection to large-scale intensity.
More Complete, More Varied Arrangements
The challenge of long-form song generation is not merely generating for a longer duration; it is creating a credible progression across sections. Emotion must build and resolve, instruments must enter, layer, and recede at the right moments, and verses, choruses, bridges, and instrumental passages must all serve a unified direction.
Music 3.0 addresses this challenge at two levels. First, section tags in the lyrics—such as [intro], [verse], [pre-chorus], [chorus], [bridge], [instrumental], [solo], and [outro]—define the song’s macrostructure. Second, the Structured Caption specifies the emotional development, instrumentation changes, vocal delivery, rhythmic foundation, ornamental timbres, and spatial effects at each stage, turning what changes, and where, into an explicit generation condition.
At the model level, Music 3.0 uses a global–local collaborative Hybrid-LM. The 8B Global LLM predicts core semantic and structural tokens frame by frame and maintains full-song context. The 0.6B Local LLM predicts acoustic tokens along the depth axis within each frame, supplying local sonic detail. This division of responsibilities maintains temporal stability across songs of up to five minutes while preserving rich variation within individual sections.
Advancing Audio Quality
Compelling composition requires equally convincing sound. Music 3.0 delivers a substantial improvement in audio quality, producing mixes that are more open, clear, and balanced, with less congestion and muddiness.
Audio modeling begins with multi-layer residual vector quantization (RVQ). The first layer uses a 16,384-entry codebook dedicated to the music’s core semantics and structure. Layers 2–8 each use a 1,024-entry codebook and progressively encode residual acoustic detail. During training, we first train the initial layer independently so that it captures core information as comprehensively as possible; all eight layers are then trained jointly. This hierarchical design balances semantic capacity, generation stability, and detail reconstruction while reducing error accumulation in long-sequence generation.
Music 3.0’s final audio rendering, however, goes beyond discrete tokens. At inference time, the synthesis stack does not load the discrete tokenizer’s decoding path. Instead, it fuses continuous hidden states from the final layers of the 8B Global LLM and 0.6B Local LLM, then generates audio directly through flow matching and the Flow-VAE. Compared with discrete tokens alone, these continuous features retain richer high-dimensional acoustic information, improving vocal pronunciation accuracy and the physical coherence of instrumental sound.
The 2.4B flow-matching module maps the fused language-model features into the VAE latent space. The 123M Flow-VAE inherits the MiniMax Speech architecture and is retrained for music-specific dynamic range and spectral distributions to reconstruct the final waveform. This enables the model to follow more precise instrumental directions and reproduce authentic performance techniques such as glissando and legato. Instrumental roles are more distinct within the arrangement, the low end retains impact without masking the mix, and fine sonic detail remains clear even in densely layered productions.
These improvements extend across the full range of musical expression: string attack and bowing in solo instruments, the impact of drums and bass, source separation in dense electronic arrangements, and the sense of space around the voice. Music 3.0 brings each of these elements closer to the listening experience of a fully produced recording.
More Natural Vocals
Vocals are often where the synthetic character of generated music is most apparent. High-frequency artifacts, rigid phrasing, unclear pronunciation, and unnatural breathing can break the listener’s emotional connection even when the song is otherwise highly polished.
Music 3.0 introduces a new audio-rendering system designed to produce more natural, studio-quality vocal performances. The Structured Caption describes vocal timbre, delivery, techniques such as breathiness and falsetto, harmony arrangement, and effects such as delay and Auto-Tune in fine detail. Continuous hidden states fused from the global and local language models carry this performance information into the flow-matching and Flow-VAE generation process. Together, these mechanisms reduce the high-frequency digital artifacts common in generated vocals while improving control over melody, pronunciation, breathing, and layered harmonies.
Vocals can respond more coherently to changes in emotion and rhythm—moving from restrained verses into expansive choruses, or from tightly articulated rhythmic delivery into sustained, open melodic lines. Harmonies unfold more naturally around the lead vocal, and breathing becomes part of the performance rather than a synthetic artifact layered around it.
From creative intent and song structure to local acoustic detail and final audio rendering, Music 3.0 advances music generation from producing plausible audio to realizing a complete, coherent creative vision. We look forward to hearing what you create.
Original source