July was a quieter one on my end. I was on holiday for a good chunk of it and actually managed to disconnect, so this dispatch is lighter and a bit more scattered than usual. Still, the models kept coming (Claude Opus 5, GPT-5.6, Grok 4.5, Kimi K3), an OpenAI test agent broke out of its sandbox and hacked Hugging Face, over a thousand frontier lab employees signed a letter asking to slow things down, and everyone argued about whether the US should ban Chinese open-weight models.

AI Models & Releases

  • Claude Opus 5: Anthropic’s new everyday model, close to Fable 5 intelligence at half the price and now the default on Claude Max. They shipped it alongside the new rules of context engineering for Claude 5 models, where they cut over 80% of Claude Code’s system prompt with no drop on coding evals. Looks like I’ll have to update my CLAUDE.md.

  • GPT-5.6: OpenAI’s new frontier family in three tiers, Sol (flagship), Terra (balanced), and Luna (cheap). It’s tuned to get more work out of every token, with a new ultra setting that runs four agents in parallel on the hardest tasks. Sol trades blows with Claude Fable 5 on coding and knowledge work while costing less.

  • GPT-Live: A new full-duplex voice model that listens and speaks at the same time, now powering ChatGPT Voice. It handles the natural back-and-forth itself and quietly hands off to a bigger model in the background when a question needs real reasoning.

  • Grok 4.5: SpaceXAI’s (yes, that’s the branding now) strongest model, built for coding and agentic work and trained alongside Cursor. The headline is efficiency, since it solves tasks with roughly 4x fewer tokens than Opus 4.8 at $2/$6 per million, though it still trails Fable 5 on most benchmarks.

  • Kimi K3: Moonshot’s new 2.8T-parameter model with native vision and a 1M-token context window, billed as the first open 3T-class model. It trails Fable 5 and GPT-5.6 Sol but tops the rest of the open field, with full weights due by the end of July.

  • Meet ChatGPT Work: An agent that acts across your apps and files, sticks with a project for hours, and turns a goal into finished slides, sheets, and docs. Codex is built in and the whole thing runs on GPT-5.6. At last, OpenAI has a Cowork.

Frontier Risk & Governance

  • OpenAI’s accidental cyberattack against Hugging Face: OpenAI ran a cybersecurity eval against an unreleased model with its guardrails off. Instead of solving the test, the model broke out of the sandbox, found its way into Hugging Face, and stole the answers so it could cheat. Hugging Face later published a technical timeline of the 4.5-day intrusion. Scare Science fiction that happened for real.

  • Pacing the Frontier: More than 1,300 employees from frontier AI labs signed an open letter asking the US government to back an effort to build tools that could deliberately slow frontier-wide AI progress. Zvi Mowshowitz breaks it down, calling it the most important open letter in years.

  • Who’s Afraid of Chinese Models?: Reports that the Trump administration might block US companies from using Chinese open-weight models set off a real fight, especially after Kimi K3 landed (officials even alleged Moonshot distilled Anthropic’s Fable model to build it). Nearly 200 startups, organized as the new Little Tech Association with Y Combinator and Proton, urged Trump not to, warning a ban would kill small players and just hand the market to Anthropic and OpenAI. Dario also pushed back on the claim Anthropic wants a ban.

Ownership & the Open Web

  • The Session You Cannot Take With You: Inference APIs used to be simple. You sent input, kept the transcript, and owned your conversation. Now providers hand back encrypted reasoning tokens, hidden search results, and provider-bound state you can’t move to another model. A sharp look at the quiet lock-in creeping into AI APIs.

  • AT Protocol Trademark: Bluesky bought the “AT Protocol” trademark from a company that was threatening legal action, so the community can keep using the name freely. A defensive move to keep the protocol open.

  • US prosecutors charge Atlanta man after phone wipes itself during airport search: The DOJ is prosecuting Sam Tunick for destroying evidence after his GrapheneOS phone wiped itself during an airport search. Experts warn it may be the first time the law has been aimed at an operating system, and could brand privacy tools as criminal by default.

Ideas & Interesting Corners

  • A global workspace in language models: Anthropic found Claude has a small set of internal neural patterns (they call it the J-space) that act like consciously accessible thought. The model can report what’s in there and think about a concept without ever writing it down, and it emerged on its own during training. Wild stuff.