GitHub HydraFusion routes one task across models, cutting cost ~65%

The Copilot CLI preview drafts, grades, and escalates across providers per task. Also: Inception's 170ms voice LLM and a free 1M-token stealth model to try.

Nowline Oct 1 11:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • HydraFusion drafts, grades, and escalates — per task

    GitHub's Project HydraFusion is a research preview that stops locking you to one model and builds a per-task plan across providers: Single runs one model directly, Cascade lets a cheap model draft and escalates to a stronger one only when a quality gate fails, and Critique adds an independent reviewer before a structured revision. On TerminalBench 2.1 it beat Claude Opus 5 by 4.9 points at 67% lower cost, and on CheckpointBench it tied on quality for 65% less. Turn it on in the Copilot CLI with `/experimental on`, pick HydraFusion from `/model`, and pay only the standard token rates of whatever models it calls — open to every Copilot tier.

  • Mercury Voice: a 170ms diffusion LLM for voice agents

    Inception's new Mercury Voice is a diffusion-based LLM tuned for real-time speech, with time-to-first-token under 170ms — fast enough that a voice agent stops feeling laggy on a live call. It's in early access at $0.40/M input and $1.50/M output, with a launch promo halving that to $0.20/$0.75, so a support bot, a live translator, or a voice game loop is a realistic weekend build.

  • A free 1M-token stealth model on OpenRouter

    OpenRouter is running an anonymous multimodal model, "Space Bunny Alpha," free during a stealth window now extended through October 5. It takes a 1M-token context and emits up to 524,288 output tokens in a single call — room to drop a whole repo or a book in and get a long, structured result back. Worth grabbing for a big-context experiment before the window closes and pricing lands.