Microsoft's MAI-Transcribe-2: $0.10/hr speech-to-text, on OpenRouter
5.2% word error across 60 languages, 10x faster than OpenAI — plus llm 0.35 ships Astra and a gated Gemini cyber model bags a 13-year-old Chrome bug.
Copy markdown
The cheapest, fastest speech-to-text — live now
MAI-Transcribe-2 posts a 5.2% average word-error rate across 60 languages (No. 1 on the FLEURS benchmark) while running ~10x faster than OpenAI's GPT-Transcribe and 7x faster than ElevenLabs' Scribe v2. At $0.10 per audio hour through year-end — a 72% cut from v1's $0.36 — it's the cheapest production STT you can wire in today via OpenRouter, Microsoft Foundry, or the MAI Playground, with diarization, word timestamps, keyword biasing, and code-switching bundled in. The catch: no real-time streaming yet, so live captioning and voice agents still wait.
Run GPT-6 Astra straight from your terminal
Simon Willison's llm 0.35 (Sept 7) adds day-one support for OpenAI's new gpt-6-astra flagship, so you can pipe it into shell scripts, cron jobs, and pytest with no SDK glue. Pair it with 0.34's llm logs --usage, which now surfaces per-response duration for quick latency and cost checks.
Gemini's gated cyber model finds a 13-year-old Chrome bug
Gemini 3.8 Flash Cyber scores 47.2% pass@1 on CWE-Bench for vulnerability discovery and reportedly surfaced a 13-year-old Chrome flaw. Access is gated behind Google's Fairwind program — request-only for now — but the bar for AI-assisted security research just moved for everyone shipping code.
Elsewhere: Anthropic reportedly locked in $517B of compute
The Information reports Anthropic has signed roughly $517B in compute contracts — at least 14.8 gigawatts — over 11 months. It's capex, not a release, but that capacity is what backs your Claude and Claude Code rate limits, and the buildout is why paid-tier headroom keeps climbing.