AssemblyAI Release Notes

The AssemblyAI MCP server lets Claude, Cursor, VS Code and other MCP clients transcribe, summarize and analyze audio on your AssemblyAI account. Ask it to "transcribe this call and list the action items."…

Summarization returned a failure status with no reason on 2 to 4 percent of requests, and on most runs for some long files. It now runs with corrected prompting and a raised output limit, so results are no longer cut…

Two more models are available through the LLM Gateway: DeepSeek V4.1 Flash 1017 (deepseek-v4.1-flash-1017, $0.30 / $1.20 per 1M prompt/completion tokens, 1M context, via Fireworks) and Reflection AI's open-weights…

GET /v2/transcript/{transcript_id} is now rate-limited to 60 requests per minute for each combination of API key and transcript ID. Requests over the limit return a 429 with a Retry-After header giving the seconds to…

Voice-activity-based silence trimming is now enabled for all English multichannel transcription, significantly reducing hallucinations caused by long silent stretches on individual channels (such as stretched filler words on a quiet channel)…

The LLM Gateway now enforces and documents per-account requests-per-minute and tokens-per-minute limits, applied per model, provider, and region over a sliding 60-second window…

Muse Spark Open is now available through the LLM Gateway as muse-spark-open. It has a 1M-token context window and supports tool calling, structured outputs, and streaming…

Google's Gemini 4 Argon is available through the LLM Gateway as gemini-4-argon: 1M-token context, $2 / $10 per 1M prompt/completion tokens, with tool calling, structured output, and streaming…

The LLM Gateway's response_format parameter (structured output) is now supported on GPT-6-family models and on Claude models including Claude Haiku 5.5. GPT-6 support shipped September 29 and Claude support October 8…

Requests made with a deleted API key previously returned "Your account is disabled", which wrongly suggested the whole account was shut off. The API now returns an error that correctly identifies the invalid key…

The legacy Summarization model (the top-level summarization boolean and summary_model parameter) and the legacy Auto Chapters model are deprecated…

Summarization now scales the number of chapters to the length of the audio, and an error where chapter timestamps could overlap has been fixed…

Universal-3.6 Pro Realtime is now live. It is a direct upgrade to Universal-3.5 Pro Realtime, trained on tens of thousands of hours of real voice-agent and telephony conversations, so the accuracy you get on quiet calls…

With speaker_labels: true, each final Turn now includes a turn-level speaker_confidence plus a per-word speaker_confidence in words[]; nothing extra to enable…

Streaming Diarization can now re-estimate speaker labels on a schedule you choose. In sessions with speaker_labels: true, set speaker_labels_revision_interval_ms and speaker labels are re-estimated over the full audio…