Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Fri 4 Sep · 14 stories · 5 sections

Frontier labs 4 items

OpenAI launches GPT-6 Astra at $10/$50 per 1M tokens, first to Daybreak-program orgs; Brockman says 'we are now in the AGI era'.
IFM releases K2 Horizon: six Apache-2.0 models from 0.9B to 375B-A23B, opening weights, training data recipes, checkpoints, and code.
Microsoft AI's MAI-Transcribe-2 speech model undercuts rivals at $0.10 per audio hour through 2026, claiming to beat Gemini 3.5.
Saudi Arabia's Humain launches Humain-M3, an Arabic model built with China's MiniMax, stirring sovereign-AI concern among US allies.

Agents & tooling 2 items

Nvidia's free PAIR beta distributes local AI inference across compatible computers on a home network, spanning RTX PCs and MacBooks.
Wired: Meta eases off 'tokenmaxxing' in employee performance reviews while pushing adoption of its internal AI agent, Hatch.

Research & papers 2 items

ARC Prize reports GPT-6 Astra at 62.7% under its own harness vs 99.9% via a provider adapter on ARC-AGI-3; Claude Opus 5 sits at 30.2%.
Google DeepMind introduces WeatherNext 3, calling it its most advanced and accurate global AI weather forecasting model.

Business & policy 4 items

Nvidia will buy Hugging Face in a $12.9B deal, second in size only to its $20B purchase of Groq's assets last December.
All G20 nations endorse the non-binding, US-proposed Carolina Principles for light-touch AI rules as tech CEOs lobby at the summit.
Wired: Cursor entered 2026 as a top-five OpenAI customer, projected to generate $1B+ in annualized revenue for OpenAI.
The Pentagon still lists Anthropic as a supply-chain risk, contradicting Lutnick's claim a day earlier that the dispute was settled.

Also notable 2 items

ChatGPT, Claude, and Grok all went down simultaneously on Thursday; the three services have since recovered.
Cerebras now serves Qwen 3.8 27B at roughly 1,500 tokens per second, per its inference model docs.
Back to top ↑