The one thing
Over the last few weeks, the two most powerful American AI models both vanished behind a government-approved rollout strategy.
The White House had OpenAI ship GPT-5.6 to about 20 vetted partners instead of the public, and the US Commerce Department finally cleared Anthropic’s Mythos to rollout again, but only for roughly 100 approved institutions. This was two full weeks after pulling it entirely.
Currently, there is a strong signal that the day-one of immediate access to frontier models are over. From a business perspective, this will be a massive advantage for approved companies.
Sources say the approved companies are part of Anthropic's Project Glasswing, which includes well-known companies like AWS, Microsoft, Google, etc. So for the foreseeable future, the playing field will not be level.
But there is still hope with Fable, which the Commerce Department indicated would be rolled out again soon. In case you missed it, both Mythos and Fable 5 use the same underlying model, but Fable 5 has a harness with guardrails for the general public.
The obvious concern is that Fable 5 is never released again. But I’d argue that a re-release with a significantly handicapped would be just as bad.
And while the US gates its best work on security grounds, the cheap and open alternative keeps shipping from China. A locked US tier and an open foreign one: that split doesn’t bode well, and begs the question, “Will America continue to dominate in AI?”
Signal
The US just put a government gate on its best AI. Mythos came back for about 100 approved institutions and GPT-5.6 shipped to roughly 20, both at Washington’s say-so, on the same day.
So what: if you build on a top US model, your access now runs through an approval list you are probably not on yet. Plan for the frontier reaching you late, or through a vendor who got cleared first.
Anthropic says Alibaba ran the biggest model theft yet: 25,000 accounts, 28.8M exchanges. Distillation: pump a frontier model for millions of answers, train a cheaper clone on them, skip the R&D bill. Anthropic’s whole argument is that this is how billions in US research becomes a cheap foreign product.
So what: it landed the same week DeepSeek made its 75% price cut permanent. Copy the capability, undercut on price, win the volume and the underlying data.
OpenAI’s first custom chip targets 50% cheaper inference. Not all the price news points up. OpenAI and Broadcom unveiled “Jalapeño,” an inference chip built to halve what it costs to serve a model.
So what: the cost of running AI is set to drop hard, and that, not export controls, is the durable American answer to cheap Chinese models. Watch this lane. It is how prices come back down for everyone.
Claude moves into Slack as a shared teammate. Not a private bot each person pokes alone. One Claude in the channel the whole team hands work to and redirects mid-task.
So what: while the strategy fights play out above your head, this is the one thing on the list you can actually deploy Monday morning.
ByteDance’s Seedance 2.5 makes 30 seconds of video in one shot. Thirty seconds of 4K from a single generation, with up to 50 reference images to hold a character steady.
So what: the best and cheapest video model right now is Chinese, and it ships in the open. The accessibility race the US just stepped back from is already being won by the other side. (Enterprise beta now, public in July.)
Noise
Skip the Google brain-drain drama, but enjoy it first, because the gossip is genuinely good. Four senior researchers walked out of DeepMind in six days, including a Nobel laureate and a co-author of the paper that kicked off this whole era, reportedly over who got the GPUs. Read it like a soap opera, not a strategy memo. Then skip the scoreboard: Anthropic’s ~$965B IPO chatter, OpenAI’s IPO timing, SpaceX buying Cursor for $60B. Big numbers, fun to watch, zero effect on your Monday.
Think the gating is overblown, or the real story of the year? Hit reply, I read every response.