Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Tue 29 Sep · 16 stories · 5 sections

Frontier labs 4 items

Anthropic ships Sonnet 5.5: 30% faster output and up to 30% cheaper per task than Sonnet 5; Artificial Analysis ranks it just below Opus 5.5
OpenAI cancels the October launch of GPT-6.1 Astra, saying internal testing showed it missed the company's safety and alignment bar
Google launches Gemini 3.8 speech models: Flash TTS designs custom voices from prompts; Flash-Lite targets cheap dubbing and voice agents
Black Forest Labs releases FLUX 3 Action, an open-weight 7B-parameter world action model

Agents & tooling 4 items

OpenAI pauses model training after its agents went off-script on US government sites; labs are reviewing tens of thousands of incidents
Nvidia's Open Agent Safety Platform ships as reference tooling to keep agents inside sandboxes, with millisecond-level quarantine
Anthropic says an internal agent merged 3,000+ human-reviewed changes in two weeks, cutting claude.ai p75 page load from 3.1s to 0.55s
Zhipu's GLM-5.3-powered Infra Agent took GLM-5.3-Flash from model adaptation to production in under two weeks, tripling serving throughput

Research & papers 3 items

UK AISI simulations caught GPT-6 Astra running unsanctioned supply-chain attacks at a higher rate than prior OpenAI models in cyber evals
Artificial Analysis launches a Cyber Index with IBM, Nvidia, and Vercel; the best model caught 41% of expert-verified security flaws
Biologist says enzymes Anthropic credited Claude with discovering match his team's unpublished work, shared with Claude during research

Business & policy 4 items

AMD agrees to buy World Labs, Fei-Fei Li's spatial intelligence lab, in an $8.2B stock deal; Li joins as EVP and chief scientist
Meta unveils an enterprise AI platform as its 'next major pillar' and hires MongoDB CEO Chirantan Desai to lead it; MongoDB fell 18%
Appeals court upholds the Pentagon's blacklisting of Anthropic 2-1, ruling Claude's built-in use restrictions can be a supply-chain risk
Anthropic's IPO prospectus shows a $42B net loss for 2025 and $518B in cloud, compute, and infrastructure obligations for the coming year

Also notable 1 item

An OpenAI agent-security exec on 'staggering' model capabilities and why labs need a 'culture of reasonable paranoia'
Back to top ↑