Morning brief

Thursday, July 2

2 people · 0 companies · 5 topics

TL;DR OpenAI proposes donating a 5% equity stake to the US government (Bloomberg, July 2); Claude Fable 5 returned globally July 1 after a 19-day export-control shutdown, the same day Anthropic shipped Claude Sonnet 5 as the new default model.

Models & research

  • model-releases Anthropic launched Claude Sonnet 5 on June 30: default for Free and Pro users, $2/$10 per Mtok intro through Aug 31 (then $3/$15). New tokenizer produces ~30% more tokens per string; simonw flagged this makes it ~1.4x more expensive for English text despite the intro pricing. Agentic benchmarks: 63.2% SWE-bench (vs Opus 4.8’s 69.2%), but 80.4% on Terminal-Bench 2.1 which actually beats Opus 4.8’s 74.6%. 1M token context on Claude Code. Anthropic · TechCrunch · MarkTechPost benchmarks

  • model-releases Fable 5 and Mythos 5 back July 1 after the US lifted export controls June 30. Anthropic trained a new cybersecurity classifier (blocks the Amazon-discovered jailbreak in >99% of cases); blocked requests now silently fall back to Opus 4.8. Pro/Max/Team users get 50% extra weekly usage through July 7 as make-good. Mythos 5 remains limited to vetted Project Glasswing orgs. Anthropic · MarkTechPost · 9to5Google

  • The 19-day Fable/Mythos shutdown forced a hard enterprise reckoning: any company that built critical workflows on a single frontier model just learned it can vanish overnight via a government order. MarketScale

  • Google released Gemini 3.1 Flash Image ($0.50/$3.00 per Mtok) and Gemini 3 Pro Image ($2.00/$12.00 per Mtok) on June 30, adding native image generation to the Gemini model family. AI-news

  • GPT-5.6 Sol is now launching on Cerebras at up to 750 tok/s; broader GA still expected mid-July. ExplainX

People

  • simonw noted June 30 that Sonnet 5’s new tokenizer means the model is materially more expensive per English character than the API pricing implies: ~1.4x versus Sonnet 4.6 for English, ~1.33x for Spanish, and roughly neutral for Simplified Mandarin. simonwillison.net

Tools & dev

  • claude-code Claude Code updated July 1: Sonnet 5 is now the default model with native 1M-token context, MCP CLI login/logout commands added, richer workflow and plugin views, and improved background-agent reliability so long-running commands survive process restarts. Releasebot

OSS

  • GitHub trending today: msitarzewski/agency-agents (2,925 stars) complete AI agency with specialized expert agents; usestrix/strix (2,167 stars) open-source AI pen-testing tool; obra/superpowers (1,047 stars) agentic skills framework; HKUDS/Vibe-Trading (918 stars) personal trading agent; JuliusBrussee/caveman (866 stars) Claude Code skill claiming 65% token reduction by simplifying prompts. GitHub trending

Other

  • sama and OpenAI proposed giving the US government a 5% equity stake in each major AI company (value ~$42.6B at OpenAI’s $852B valuation). Structure: donation rather than sale, framed like the Alaska Permanent Fund returning AI dividends to citizens. Whether Anthropic, Google, and Meta would accept is unclear. Bloomberg · CNBC

  • California Governor Newsom signed a deal making Claude the first AI productivity tool available to all state agencies, cities, and counties. AI-news roundup

  • The UN’s Independent International Scientific Panel on AI (co-chaired by Yoshua Bengio and Maria Ressa) released its preliminary report July 1; findings will be presented at the Global Dialogue on AI Governance in Geneva July 6-7. UN roundup