Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Thu 8 Oct · 16 stories · 5 sections

Frontier labs 3 items

Anthropic launched Claude Haiku 5.5, the first Haiku with effort controls, at $0.10/$0.50 per M tokens under 100K - GPT-6 Luna's price.
OpenAI rolled GPT-6 out to every ChatGPT tier with Intelligent UI: charts, buttons, and forms in answers; Sol for paid, Luna for free.
Reflection AI released Beam, a 501B-param (23B active) Apache-2.0 MoE it says matches GLM-5.2 on reasoning at 3-4x less inference compute.

Agents & tooling 4 items

Anthropic shipped a beta Claude sidebar across Google Docs, Sheets and Slides (paid plans): it edits files directly or asks per change.
Hark, Brett Adcock's startup, launched Hark Pro: a proactive assistant whose Handoff cloud computer drives up to 36 browsers per task.
Docker released docker-agent, a CLI plugin that builds and runs AI agents from declarative YAML, with tools and multi-agent orchestration.
Voice-AI startup Sesame introduced Personal Agents that get their own computer and converse by text or speech; AI glasses follow in 2027.

Research & papers 3 items

Vals AI ran Claude Opus 5.5 agents over physics simulations hunting next-gen memory materials, surfacing two room-temperature candidates.
NVIDIA fine-tuned one Nemotron model family to gold-level results on both IOI and IMO and published the recipe on Hugging Face.
Mathematicians push back on OpenAI's 722-paper drop: Tao urges valuing progress 'holistically'; Aaronson weighs a 'Mathocalypse'.

Business & policy 4 items

SemiAnalysis measured API-equivalent value of AI subscriptions: Claude plans deliver ~5x OpenAI's per dollar on agentic workloads.
Anthropic halved Sonnet 5.5 cache-read pricing to $0.10/M and added monthly API credits: $100 on Max 5x, $200 on Max 20x, $500 team cap.
Nous Research raised $90M to bring Hermes, its open-source agent with 22M+ downloads since February, to enterprises.
OpenAI will watermark ChatGPT and Codex text in the EU under the AI Act; opt-in elsewhere, and 25% synonym edits cut detection to 17%.

Also notable 2 items

Google launched SynthID Detector, a site that checks whether images, video, or audio came from Google, NVIDIA, OpenAI, or Kakao models.
Microsoft opened $5,999 preorders for the Surface RTX Spark dev box: 128GB unified memory for 120B+ local models, shipping in November.
Back to top ↑