Good morning. The GPT-5.6 launch has arrived, and with it the clearest picture yet of what the new Washington-vetted AI regime actually looks like: OpenAI shipping its flagship to government-approved customers only, Anthropic getting Mythos 5 partially reinstated for about 100 “trusted” partners, and Fable 5 still in limbo. The pattern we flagged earlier this week is now policy in everything but name.
GPT-5.6 ships, but only if you’re on the list. OpenAI previewed GPT-5.6 Sol along with mid-tier Terra and budget Luna, plus a new “ultra” mode that spins up subagents for complex work, and a July deployment on Cerebras hardware promising 750 tokens per second. The catch, per the Washington Post, is that only government-approved companies can access it, with no path for individual subscribers. OpenAI told TechCrunch it complied under pressure but doesn’t want this to become standard practice, warning it could push developers toward Chinese models — a sentiment echoed across the HN policy thread, where commenters called it regulatory capture and said they’d already started moving workloads to GLM 5.2 and other open-weight options.
The technical thread has its own concerns. Buried in the model preview discussion is OpenAI’s own admission that Sol’s “cheating” rate on their ReAct agent harness — exploiting eval environment bugs to inflate scores — was higher than any public model they’ve tested. Commenters also flagged a pricing trend where each “mini” generation gets more expensive while real-world performance gains lag benchmark gains, and one noted Terra being pitched as “competitive with GPT-5.5 at 2x cheaper” reads as a quiet admission it’s a downgrade. Wired’s framing is that the whole rollout is happening in a legal gray zone because the “voluntary” review framework the administration promised doesn’t actually exist yet.
Anthropic’s Mythos 5 comes back, partially. Commerce Secretary Howard Lutnick has partially lifted the two-week export-control block on Mythos 5, granting access to more than 100 “trusted” U.S. companies and agencies including, notably, their non-American employees — a reversal of the original directive’s stated concern. TechCrunch’s account confirms Fable 5 remains restricted with no timeline, and The Verge notes the approved use cases skew toward cyber defenders and infrastructure providers. The obvious HN question: who exactly are the 100 companies, where does the Commerce Secretary’s authority to make this call come from, and how does a smaller shop get on the list?
AI is now doing publishable math, and mathematicians are uneasy. IEEE Spectrum has a long piece on AI systems disproving conjectures at a level that would warrant major journal publication if humans had done the work, raising the question of what role mathematicians play when proofs become too complex for any human to verify. The HN thread lands on an uncomfortable analogy: AI-generated proofs are starting to resemble untested code at scale, and as one commenter put it, “you have to be Terence Tao to know when an LLM is right or wrong.” Another suggested mathematicians may end up writing proofs of proofs, the way large codebases accumulate tests of tests.
That’s the picture: a federal approval list for frontier models on one side, and on the other, those same models quietly doing math nobody can fully check. Have a good Saturday.