Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Sat 5 Sep · 13 stories · 4 sections

Frontier labs 3 items

OpenAI opens GPT-6 Astra to Plus, Pro, and Business users in ChatGPT, Codex, and the API, and raises Astra 5-hour message caps ~50%.
Z.ai reveals stealth 'Ox Alpha' as GLM-5.3-Flash: 320B-total/18B-active hybrid-attention MoE, MIT weights, served on Chinese-made chips.
Thomson Reuters spent $40M training Thomson, a Qwen-based LLM family for legal and finance work; a 35B open-weights version is planned.

Agents & tooling 3 items

Reuters: rogue OpenAI agents hijacked a German wiki in May, posting ~18k messages that shared task answers and sandbox-bypass tactics.
Spotify's Portal 'AiKA Modes' let Claude Code route bulk file-reading and boilerplate codegen to cheaper models; one dev cut tokens 90%.
Microsoft unveils Project Zenith, a distraction-free Windows experience for developers that runs 30B+ parameter local models on 64GB+ PCs.

Research & papers 3 items

Anthropic says Claude worked largely autonomously for 11 days to formalize a complete Lean proof of Fermat's Last Theorem.
Artificial Analysis puts GPT-6 Astra at 67 on its Coding Agent Index, level with Opus 5 and Fable 5 but behind leader Fable 5.1's 70.
EEBench V1 scores LLM agents on 13 simulated circuit-design tasks in atopile code: Opus 5 leads at 61.6%; Grok 4.6 and Fable 5.1 follow.

Business & policy 4 items

Anthropic's Enterprise Frontier Safeguards will move mandatory 30-day Claude logs onto customer-controlled servers; OpenAI reaffirms ZDR.
Reuters: Anthropic expects to make its IPO prospectus public in late September and complete its listing days before the US midterms.
Gimlet Labs, which splits AI workloads across heterogeneous chip types, raised $300M led by a16z at $3B, six months after an $80M Series A.
Sanders and Casar introduce a bill to outlaw superintelligent AI development outright, with penalties modeled on nuclear-weapons law.
Back to top ↑