ChatGPT Voice takes the inbox, and Muse's files match OpenClaw

📰 News

• Talk, and it handles the inbox — ChatGPT Voice now runs on GPT-6 Astra, Sol, and Luna, and it can open email, calendar, and Slack. Check a meeting, send mail, or cancel a duplicate charge, all by voice. Only Plus and Pro get the phone Work tab. → The Decoder · TechCrunch

• Muse: half a million in a week — US app and WhatsApp. 250,000 people a day, more than two million prompts. Bookings, expenses, calendars. Files like SOUL.md nearly match OpenClaw. Heavily inspired, built from scratch, and Friedman’s answer on the copies: Peter got those things exactly right. → The Decoder

• Smarter, worse at prose — Kernion fine-tunes Claude. Lean the rewards toward math and code, and the model writes dense explainers for other models, not for people. Opus 4.6 was the last writing model. 5.5 is the best since, and still not past 4.6. → The Decoder

• A voice from a sentence — Gemini 3.8 Flash TTS. Describe the role, accent, and character. 2,000 presets, a 30-second clone (they have to consent), SynthID watermark. Cheaper Flash-Lite is for dubbing and voice agents. An hour this year: Flash $0.81, Lite $0.54. → The Decoder

• Gemini in the edit — Studio reads the script and the rough cut and flags the pacing. A/B three cuts. Each viewer gets one of three thumbnails. Shorts trims and lines up the music. 72% of US creators aged 14 to 44 already use AI to make or edit. → The Decoder

• Missing from the prospectus — Nvidia-backed Nscale. ByteDance was 73% of $33 million in 2025 revenue, and the S-1 body never names them. 2,304 B200s in Norway. They expect that share under 20% this year. → The Decoder

🔥 GitHub Trending

• google/ax — Declare an agent task. It attaches a sandbox, a repo, and the hosts that task is allowed to call.

• agent-substrate/substrate — Packs a lot of those sandboxes. Wakes an idle agent in under half a second. Not an official Google product.

• browser-use/video-use — Hand a footage folder to a coding agent, get final.mp4 back.

• superdesigndev/treg — OpenRouter for tools. One token, 3,000+ calls, priced per call.

• davila7/claude-code-templates — Install Claude Code agents, slash commands, tool connectors, and hooks.

▶️ YouTube

• Anthropic went CRAZY (Opus 5.5) — Matthew Berman. Opus 5.5 tests, with Thariq.

• GPT-6 Luna Might Be the Best Deal in AI — Matthew Berman. Whether Luna is the cheap model to actually use.

💬 Community

• A TTS playground — Simon Willison built a bring-your-own-key page. 1 minute 18 seconds of audio, 2.74 cents.

• llm 0.36 — The terminal LLM CLI adds gpt-6-sol and gpt-6-luna. One-turn models can refuse a chat history.

• A year earlier — Decision models that don’t emit one token at a time. The author says a frontier lab later called that a breakthrough.

• The Jalapeño chip — IEEE Spectrum on OpenAI using its own LLMs to design the chip.

• October 14, San Francisco — Simon Willison and Jesse Vincent. A show-and-tell for coding-agent experiments. Not pitches.