ElevenLabs Doubles to $22B on Voice Agents, Microsoft Ships Streaming Speech-to-Text, and GitHub Copilot Keeps Retiring Models Under Its Users

Voice agents are attracting serious money and tooling, while the model layer underneath every chatbot keeps shifting. This week's stories share one theme: the parts of your stack you do not control are changing quickly. We could not open primary pages from our environment, so details come from search summaries and secondary coverage, and we flag where sources disagree or figures are vendor claims.

1. ElevenLabs Doubles Its Valuation on Conversational Agents

ElevenLabs announced on September 30 that a $300 million employee tender offer closed at a $22 billion valuation, double the $11 billion set in its February Series D. One report gives the amount in pounds rather than dollars, so treat the figure as approximate. A tender sells existing shares, so the company receives no fresh capital from it.

What makes this relevant to builders is the stated driver. Coverage ties the valuation to ElevenAgents, its platform for voice and chat agents, reporting more than triple the ARR since February, enterprise customers at about 55 percent of revenue, and over 15 million conversations a week. Those numbers are secondhand and unaudited, but the direction matches what we see elsewhere: voice is moving from demo to line item.

2. Streaming Speech-to-Text Lets Agents Think Mid-Sentence

On October 1, Microsoft AI released MAI-Transcribe-2-Streaming, its first streaming speech-to-text model, alongside MAI-Voice-2.1 and MAI-Voice-2.1-Flash text-to-speech models, all through Microsoft Foundry. Microsoft says it supports 60 languages with continuous language detection and returns first partial hypotheses in just over 100 milliseconds, then revises them. The claimed 2.5% final word error rate is a vendor figure, and one tracker notes the model had not yet appeared on the public leaderboard it cites. Azure's documentation marks it a public preview with no SLA and not recommended for production. Introductory pricing is reported at $0.54 per hour of audio through the end of 2026.

The architectural point matters more than the benchmark. Partial transcripts arriving over a WebSocket mean an agent can start retrieval or a tool call before the speaker finishes. The catch is that partials get revised, so anything you launch early must be cheap to cancel:

let inflight: AbortController | null = null;

stt.on('partial', (text) => {
  inflight?.abort();                 // the hypothesis changed; drop stale work
  inflight = new AbortController();
  prefetchContext(text, inflight.signal); // read-only, side-effect-free
});

stt.on('final', (text) => respond(text)); // only act on finals

Keep speculative work read-only. Writes, payments and messages should wait for the final transcript.

3. Model Retirement Is Now a Monthly Event

GitHub has deprecated Copilot models in successive rounds: six on September 1, including Claude Opus 4.5 and 4.6, then on October 2 Gemini 3.5 Flash, Gemini 3.6 Flash, Kimi K2.7 Code and Claude Opus 4.7. One detail shows how fluid this is: GitHub's September 16 notice and its October 2 changelog name different replacements for Opus 4.7 (Claude Opus 5 versus Opus 5.5). We also saw a newsletter claim of further removals on November 16, but could not find it in GitHub's own changelog, so treat that as unconfirmed.

Whatever the exact dates, the lesson for anyone shipping a chatbot is that a model id is a dependency with an expiry date. Keep ids in configuration rather than scattered through code, and run a small answer-quality suite before switching:

const MODEL = process.env.CHAT_MODEL ?? 'primary-model-id';
const FALLBACK = process.env.CHAT_MODEL_FALLBACK; // tested, not just configured

A fallback you never exercise is a guess. Run your evals against both ids on a schedule so a retirement notice becomes a config change instead of an incident.

Conclusion

Capital is flowing into voice, streaming components are making it faster, and the models beneath all of it are being swapped on someone else's calendar. Design for interruption in the audio path, keep speculative work cancelable, and treat model ids as expiring dependencies. Verify vendor numbers against primary sources before you plan around them.

A latency timeline showing partial transcripts, speculative retrieval and the final response would suit this post well.

— Maya