Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Wed 23 Sep · 16 stories · 6 sections

Frontier labs 3 items

Anthropic launches Claude Opus 5.5 at $4/$20 per 1M tokens, a 20% cut, claiming Fable 5.1 parity on most tasks plus stronger safeguards.
OpenAI launches GPT-6 Sol ($2/$10 per 1M tokens) and Luna ($0.10/$0.50), both about half the price of their GPT-5.6 equivalents.
OpenAI ships improved GPT-6 prompt caching: higher default hit rates, a 30-minute reuse window, and new dashboard plus miss diagnostics.

Agents & tooling 4 items

HarnessRouter open-sources a unified local interface to 14 agent harnesses (Codex, Claude Code, Jev) under one task protocol and API.
Meta launches Muse connectors so outside apps can plug into its AI agent, with a directory of approved integrations for developers.
Nat Friedman says Meta built Muse 'from scratch' but was 'heavily inspired' by OpenClaw; internal data shows 500K users in week one.
Rabbit launches OS3, a cloud agent that drives local apps and files on Windows, Mac, and Linux via web, Telegram, iMessage, or its R1.

Research & papers 3 items

Study maps a 'pain axis' inside 25 open models that spikes under mistreatment; amplifying it made some pick user-harming relief options.
GPT-6 Astra autonomously broke a 1941 Enigma message unsolved since 2005, writing its own simulator and Bombe, Crypto Cellar validates.
Artificial Analysis ranks Xiaomi's MiMo-v2.6-Pro the top open-weight model on its Intelligence Index, tying Grok 4.7 and beating GLM-5.3.

Business & policy 4 items

Sources: China's cyberspace regulator is probing DeepSeek and Moonshot over alleged data leaks to Anthropic via Claude, questioning staff.
At the UN General Assembly, Trump says the US is 'officially' renaming AI 'super intelligence' and rejects any global AI governance scheme.
OpenAI says it will let third-party groups run technical safety evaluations of its models during training, evaluation, and deployment.
Alibaba plans a 5T-10T-parameter model and unveils the Zhenwu V900 accelerator, claiming 3x prior performance and 500K-unit clusters.

Tutorials 1 item

Hands-on: rebuild Jev-style fixed-answer scoring locally with SGLang's /v1/score, tested on Qwen and DeepSeek against text generation.

Also notable 1 item

A researcher asked Meta's Muse to archive its files and got a 6.8GB runtime filesystem, with internal docs, skills, and SSH key files.
Back to top ↑