AI's July 3 shake-up: rules, models, science

AI just hit a new phase: governments are shaping releases, frontier models are bouncing back online, and Sonnet 5 is already moving the coding bar. Here are 4 updates worth your attention^[1][5].

White House frontier AI talks: the U.S. is close to voluntary rules with OpenAI, Google, and Anthropic on benchmarks, testing timelines, and access rules. OpenAI has even floated a 5% U.S. equity stake^[1][5].

Fable 5 is back worldwide after a 19-day export-control pause. Anthropic says the new classifier blocks the jailbreak technique in 99%+ of attempts, but some security prompts now get routed to Opus 4.8. July 7 is the billing cliff^[5].

Claude Sonnet 5 is now the default Free and Pro model, with a 1M-token context window and a 63.2% agentic coding score. Early users say it handles multi-step work better, though tokenizer changes may raise costs^[1][5].

Anthropic’s AI for Science event says deterministic tools can lift biology accuracy from 16.9% to 92.8%, and it opened grants for up to 50 projects with up to $30k in Claude API credits each. Science is getting a real AI workflow boost^[1][5].