The top tier of consumer AI has been moved out of reach. In a single news cycle, OpenAI launched its strongest GPT-5.6 model and Anthropic returned its Claude Mythos 5 — and in both cases access was limited to a list of organizations approved by the US government. If you are choosing an AI assistant or model tier right now, the decision is no longer about which frontier model leads on benchmarks. It is about which capable model you are actually allowed to buy and run.
What was restricted, and how
OpenAI's GPT-5.6 family comes in three variants, according to Engadget, spanning its most powerful and its most affordable models to date. The most powerful variant did not ship to the public. The Next Web reports it went to roughly 20 partners individually approved by the US government, and The Decoder adds that OpenAI is releasing it on a "customer by customer" basis at the government's request.
Anthropic faced a similar constraint. 9to5Mac reports that a directive forced the company to disable both Mythos 5 and Fable 5 for all customers, before it was cleared to release Mythos 5 to a select group of US organizations. Frandroid confirms Mythos 5 is back for that limited group, while Fable 5 remains blocked.
This is not presented as permanent. Gizmodo reports that OpenAI does not believe this kind of government access process should become the long-term default. For a buying decision today, the stated intent matters less than the current state: the flagship tier is gated, and most buyers cannot reach it.
Why this changes the comparison
For individuals and small teams, frontier benchmark scores were already of limited use. A model you cannot subscribe to does not belong on your shortlist. The relevant field is now the widely available tiers: the mid-range and cheaper variants of GPT-5.6, the still-accessible Claude models, and the open-source alternatives that carry no approval requirement at all. Peak capability tends to show at the edges of difficulty, not in the drafting, summarising and coding assistance that fill most subscribers' days.
The widely available tiers
OpenAI's three-variant structure means the most affordable GPT-5.6 model is generally accessible, and Engadget describes it as the cheapest the company has shipped. For high-volume, cost-sensitive work, price-per-task on a mid or low tier often matters more than peak capability. Anthropic's accessible Claude tiers remain a credible option for users who prefer its output style, with the caveat that the lineup has just been disrupted and Fable 5 is unavailable.
The open-source case
Open and self-hostable models sit outside any approval process. The Decoder reports that the AI startup Lindy switched entirely from Claude to Deepseek, with its CEO citing cost pressure as a matter of survival. That decision was about economics rather than access, but it points to the same conclusion: for teams running at scale, a model you control on price and availability can beat a more capable one you need a vendor, or a government list, to grant.
Two caveats before you commit
First, availability is now a moving target. A model that is accessible this month may be restricted next month, and vice versa. Avoid building a workflow that depends on a single gated tier, and keep a fallback model you can switch to without re-engineering everything.
Second, output is not neutral. The Decoder reports that most major chatbots still lean left on political questions, with one investigation finding GPT-5.5 gave exclusively left-leaning arguments 80 percent of the time. If your use case touches contested subjects, test the specific model you intend to deploy rather than trusting a benchmark or a brand.
What to do now
- Ignore the gated flagships. If you cannot subscribe to it, it is not a product you are choosing between. Compare the tiers actually on sale.
- Match the tier to the task. For routine work, a mid or cheap variant usually delivers most of the value at a fraction of the cost.
- Keep an exit. Favour setups that let you change models. Open-source options reduce exposure to both pricing and access changes.
- Test for your own use case, including any bias that affects your domain, before you standardise on one model.
The headline this cycle is that the best models are locked away. The more useful point is that, for most buyers, they were never the right choice anyway. Judge what you can actually run, and what it still does for you months from now. For ongoing coverage, see our AI news, the AI assistants comparison, our top AI assistant picks, and the glossary for the terms above.