Eleven Realtime: ElevenLabs' speech-to-speech agents at ~570ms
Drop it into OpenAI Realtime code by swapping the endpoint. Plus Cloudflare's token-free Clairic, open model SolarRenderer, and Anthropic's kill-switch bill.
Copy markdown
Listen, reason, speak — in one model
ElevenLabs shipped Eleven Realtime, a speech-to-speech voice-agent family in two sizes: Max reaches ~570ms to first audio and 98.8% on Big Bench Audio, Mini ~620ms at lower cost. It handles turn-taking, interruptions and tool calls across 90+ languages, so you can drop a full voice loop into an app without wiring STT→LLM→TTS yourself.
A drop-in for OpenAI's Realtime API — with strings
It speaks OpenAI's Realtime protocol over WebSocket, so existing apps just point at wss://api.elevenlabs.io/v1/convai/realtime with model eleven_realtime_v1_mini. The catch: WebSocket-only (WebRTC apps re-plumb transport and handle 24kHz PCM16 audio themselves), you can't swap the underlying models, and Mini lacks the Zero-Retention/HIPAA support Max carries in the US and EU.
Cloudflare's Clairic returns values, not tokens
Clairic, a new 'System One' extraction model, is in preview on Workers AI as @cf/cloudflare/clairic — it returns typed fields (text spans with offsets and confidence, computed dates/numbers, bounded estimates, even short forecasts) in a single pass with no reasoning tokens. In Cloudflare's own claims pipeline it ran 3.2x faster (19.3s vs 61.7s) and 49% cheaper ($6,032 vs $11,740 per million claims); open weights are promised.
SolarRenderer: open, controllable video, fast
A new open release pairs a frozen 33B backbone with a single control LoRA to render video from scene structure, reporting a 12.3x speedup (209s vs 2,562s for a 158-frame clip) and up to 28 FPS at 768p in its block-causal variant. Weights, code and dataset are public — a weekend path to layout-driven, controllable video without training your own model.
Anthropic floats a federal model kill-switch
A draft bill from Anthropic would let the Commerce Department ask a federal judge to halt a model's release if it poses an imminent risk of more than 50 deaths or over $1B in damage. It's early, but if it moves it shapes which frontier models ship and when — the upstream of what you'll eventually get to build on.