OpenAI restricted GPT 5.6 general access at Washington's request. Anthropic pulled Mythos 5 and Fable 5 offline under a government directive. Both landed in the same month. The mechanism is classification, not safety review.
The headline numbers around OpenAI's new GPT-5.6 model family are the benchmarks and the price. The consequential story is that the US government is now deciding who gets to use the model at all.
OpenAI released GPT-5.6 as a three-tier lineup on June 26, 2026: Terra for everyday tasks, Luna for speed and efficiency, and Sol as the flagship, with a higher-end Sol Ultra that orchestrates smaller sub-models for harder problems. Heise.de's write-up of the announcement framed the launch around performance per token, an efficiency story for developers paying per query. The model did not arrive as a normal consumer release. General-public availability was constrained at the request of the US government, according to OpenAI's own framing to TechCrunch and a separate heise.de report on the restrictions.
That pullback is the real news, and it is not new. Three weeks earlier, the US government had ordered Anthropic to suspend access to its most capable models, Mythos 5 and Fable 5, according to Anthropic's own statement. Anthropic called that step a "directive." OpenAI is now describing an equivalent constraint on Sol as a "request." Both companies are withholding the same category of frontier model from the general public in the same month, and both are crediting Washington's discretion for the gap.
The mechanism behind the pullbacks is classification, not safety review. The decision-maker in this story is the US government's determination that these models qualify as controlled dual-use technology, not a finding by an outside lab that they are unsafe. The companies complied. Anthropic's earlier safety warnings, which had publicly stressed the cyber and bio risks of its frontier models, did not keep Mythos 5 and Fable 5 available; if anything, they appear to have made the models easier to classify as restricted. OpenAI's safety pitch on GPT-5.6, including improved jailbreak resistance and real-time cyber-misuse classifiers per the deployment safety system card, is not presented as the reason the model is restricted. It is presented as the reason the restriction should be temporary.
The pricing still matters, because it shows where the competitive pressure has settled. OpenAI priced GPT-5.6 at $2.50 per million input tokens and $30 per million output tokens, undercutting the $10 input and $50 output that Anthropic was charging for Fable 5 and Mythos 5, per the same heise.de coverage. On OpenAI's Terminal-Bench 2.1 coding results, the company reports Sol at 91.9% and Sol Ultra at 88.8%, with the comparison model "Mythos" listed at 88%. The naming is worth pausing on: OpenAI's competitive referent is a model Anthropic has not publicly announced under that exact name, and the heise summary uses "Mythos 5," "Fable 5," and "Mythos Preview" interchangeably without identifying which Anthropic system is actually being measured. Those benchmark numbers are OpenAI-reported, not independently reproduced, and Sol Ultra scoring lower than Sol on the same chart is an internal contradiction the announcement does not flag.
The regulatory backdrop makes the picture more confused, not less. In June 2026, the White House signed Executive Order 14409, "Promoting Advanced AI Innovation and Security", which on its face pushes for faster AI deployment and tighter cyber-defenses. A Wiley Rein analysis of the order reads it as accelerating frontier AI rollout. Two of the largest US frontier labs are simultaneously being told to keep their top models off the general-public shelf. The executive order and the pullback are not formally contradictory, but they sit in obvious tension, and OpenAI's framing leans on that tension by publicly arguing that the GPT-5.6 restriction "should not be the norm."
The practical effect is that the boundary between "developer-accessible" and "publicly usable" AI is now drawn in Washington, not in Mountain View or San Francisco. A model can be cheap, fast, and benchmark-leading, and still be unavailable to a curious user with a credit card. The frontier AI race in 2026 is being run inside a classification regime that neither lab chose and neither lab can unilaterally lift.
What to watch next is whether the US government's June actions harden into a standing gate for the top capability tier, or relax as OpenAI and Anthropic ship the safety documentation each side says is coming. The first new Mythos- or Sol-class model that reopens to the general public, and the paperwork that lets it, will tell readers whether Washington's June intervention was a one-off or the new floor.