ElevenLabs Scribe v3 Is Live: Three API Changes That Matter

ElevenLabs released Scribe v3, a speech-to-text model available immediately through the existing Speech-to-Text API, which delivers 27% fewer word errors than Scribe v2 across 102 languages on the FLEURS benchmark and reaches 1.9% AA-WER, tying AssemblyAI for the top of the Artificial Analysis STT leaderboard. The company also introduced Workflows as Code in its Python and JavaScript SDKs from v2.77.0, letting developers version-control agent conversation flows, and set a March 1, 2027 removal date for the deprecated `eleven_text_to_sound_v2` model, which will break integrations that pass that model ID explicitly. The real-time `scribe_v3_live` variant adds speaker labeling, turn prediction and a keyterm library of up to 10,000 terms at $0.22/hr, and is now the default transcription layer inside ElevenAgents.

ElevenLabs held its Summit NYC at Lincoln Center last week and dropped 13 product announcements. For most developers, the number is a distraction. What matters is narrower: Scribe v3 is live in the API, agent workflows can now be version-controlled in code, and one deprecated model has a hard removal date that will break integrations silently if you don’t act. Scribe v3: Ship It This Week The headline release is Scribe v3, ElevenLabs’ new speech-to-text model available immediately via the existing Speech-to-Text API. It delivers 27% fewer word errors than Scribe v2 across 102 languages on the FLEURS benchmark, reaching 1.9% AA-WER — joint top of the Artificial Analysis STT leaderboard https://artificialanalysis.ai/speech-to-text/non-streaming alongside AssemblyAI. Switching costs nothing beyond a one-line change to your model id : curl -X POST https://api.elevenlabs.io/v1/speech-to-text \ -H "xi-api-key: YOUR API KEY" \ -F "model id=scribe v3" \ -F "file=@audio.mp3" \ -F "language code=en" The real-time variant, scribe v3 live , adds speaker labeling mid-conversation, turn prediction, and a keyterm library supporting up to 10,000 domain-specific terms. It’s priced at $0.22/hr and is now the default transcription layer inside ElevenAgents — meaning any voice agent already on the platform gets the accuracy upgrade automatically. For mixed-language audio, use secondary language codes . For sensitive pipelines, set enable logging: false for zero-retention mode. Workflows as Code: Version-Control Your Agent Logic The less-hyped but more significant shift is Workflows as Code, available in Python and JavaScript SDKs from v2.77.0. You can now build, version, and deploy agent conversation flows entirely in code — conditional routing, tool call triggers, multi-agent handoffs, all serializable to JSON for storage in your repository: triage.when Expr.variable "tier" .eq "gold" .transfer "vip", Destination.agent "agent vip" Serialize for version control json str = wf.to json string Workflows created in code remain editable in the visual editor — no lock-in. For teams already running CI/CD on AI pipelines, this closes a real gap. ElevenLabs also added 24 new tools to its Hosted MCP Server https://elevenlabs.io/docs/eleven-agents/customization/agent-workflows-as-code for agent inspection: analytics, configuration, alerts, and test invocation. The Urgent One: SFX v2 Is Removed March 1, 2027 The item most likely to generate a 3 AM incident: eleven text to sound v2 is deprecated and will be removed on March 1, 2027. If your code passes model id: "eleven text to sound v2" explicitly, change it to "eleven text to sound v3" before that date. If you don’t specify a model id , the platform auto-migrates and you’re clear. The risk is code that locked the model ID months ago and hasn’t been touched since. GitHub issues are already appearing across open-source projects. Run a codebase search, check the November 2 changelog https://elevenlabs.io/docs/changelog/2026/11/2 , and migrate now rather than under pressure in February. SFX v3 handles multi-part prompts and vocal or musical sounds more accurately anyway — there’s real upside beyond avoiding breakage. ElevenAI the App: Useful Preview, Not an API Yet The Summit’s marquee announcement was ElevenAI, a team app combining meeting transcription, cross-meeting search, dictation, and ElevenAgent-driven follow-up actions. It runs on Scribe v3 internally, supports 90+ languages, and targets the Otter.ai and Fireflies market. It’s in free early access — sign up at elevenlabs.io, don’t look for an app download. There’s no developer API surface yet. When it opens one, the combination of real-time transcription and agent-driven actions will matter. For now, it’s a useful preview of platform direction. What to Do Now - Update any model id calls to scribe v3 for immediate accuracy gains - Upgrade to Python or JS SDK v2.77.0+ to access Workflows as Code - Search your codebase for eleven text to sound v2 and migrate to v3 - Check the Scribe v3 API reference https://elevenlabs.io/docs/api-reference/speech-to-text/convert for the full parameter list ElevenLabs shipped 550+ partners, announced 15 MW of NVIDIA Vera Rubin compute, and previewed enterprise custom model fine-tuning. The Workflows as Code release and the ElevenAgents Unity SDK for spatial audio signal where the company is heading: infrastructure, not just an API. The free tier still runs up to the Creator plan at $22/month https://elevenlabs.io/pricing for most self-serve builders.