Washington decided this week that it gets a say in which AI models ship. OpenAI previewed GPT-5.6, a three-model family called Sol, Terra, and Luna, but at the White House’s request the rollout stays limited to trusted partners for now, close to the script Anthropic followed with Fable 5 a few weeks ago. That is starting to look like the normal way frontier models launch, not the exception.
Nathan Lambert thinks that pattern will matter less than it looks. His latest read on the open ecosystem argues that open models now come from pure model-makers, big tech, and product companies at once, not just a couple of labs in one country. I think he has a point. Gatekeeping one company’s release only slows the field down if nobody else is shipping weights that same week.
Anthropic had a full week of its own. It shipped Claude Sonnet 5, its most agentic Sonnet yet, matching near-Opus performance on coding and tool use at a lower price, and made it the default model across every plan. It also redeployed Claude Fable 5 and Mythos 5 worldwide now that export controls lifted, with a stronger safety classifier and a new cross-lab framework for scoring jailbreak severity, built together with Amazon, Microsoft, and Google.
Simon Willison flagged a German court ruling that holds Google liable for errors in its AI Overviews, treating an AI agent as the legal agent of whoever deployed it. Small ruling, big implications if other courts pick it up. Jack Clark’s Import AI 463 covered a Chinese cluster now running more than 10,000 GPUs and a system built to let robots improve themselves with less human tuning. A new dataset called TraceLab opened up 4,300 real Claude Code and Codex sessions, showing that coding agents run long, messy loops that today’s serving infrastructure was not really built for.
Six months ago, a model release was a company decision. Now it is half company, half government, and the open weight crowd is betting that will not matter much. Worth watching which bet pays off.
T.
