The US gated its two newest flagships in a day. The models you can still call are open weights.
On June 26, OpenAI held GPT-5.6 back to government-approved partners and Commerce let Anthropic re-open Claude Mythos 5 to a short list of trusted US organizations. If you build outside that list, the two newest American frontier models are not yours to call. The good news for your budget: the field you can still reach is cheaper, and a lot of it ships its weights.
Photo by Tim Umphreys on Unsplash
- Two American labs gated a flagship on the same day, June 26. OpenAI held
gpt-5.6(Sol, Terra, Luna) to government-approved partners; Anthropic was cleared to bringclaude-mythos-5back for a small set of trusted US orgs only, after suspending it June 12. - Both run through the June 2 executive order, which lets the federal government take up to 30 days of pre-release access to "covered frontier models." The order is voluntary on its face, the two actions used different legal mechanisms, and neither names a country.
- If you build outside the approved circle, the two newest US flagships are not yours to call. The field you still can runs from $0.14 per million tokens up: DeepSeek, GLM-5.2, Kimi, Qwen, Mistral, MiniMax, Doubao.
- Most of the cheap rows ship open weights, so no government can switch them off. That is the part of the table that actually matters this week.
What changed on June 26
OpenAI previewed GPT-5.6 to a small group of trusted partners in Codex and the API, partners whose participation was shared with the government, with general availability promised in coming weeks. It is the first time OpenAI has staggered a flagship launch at Washington's request, and the company said plainly that it does not think this kind of access process should become the default. The same day, Commerce cleared Anthropic to re-open Mythos 5 to a limited group of trusted US organizations that defend critical infrastructure, two weeks after the model went dark.
Both moves run through the June 2 executive order on advanced AI. It defines a "covered frontier model" using a classified cyber-capability benchmark the NSA Director maintains, and it lets a developer give the federal government up to 30 days of access before other trusted partners get in. We covered the first shock of this when Fable 5 and Mythos 5 went dark on June 12. What is new is that this is now a pattern, not a one-off: the newest model from each of the two largest US labs is, today, something most of the planet cannot use.
Two different locks on the same door
It is tempting to read "the US gated two models" as one event. It is two, and the difference matters if you are deciding how much to trust the rest of an American lab's lineup. The executive order itself is voluntary on its face: it says nothing in it authorizes a mandatory licensing or preclearance requirement for new models. OpenAI's GPT-5.6 hold reads as exactly that, a voluntary delay it agreed to and openly grumbled about.
Anthropic's June 12 suspension was described in harder terms, an export-control directive that ordered the company to cut off Fable 5 and Mythos 5 for any foreign national, including its own foreign-national staff. That is a stronger legal instrument than a request you can decline. The June 26 re-opening softened it back to a narrow allowlist. The count of organizations on that list has been reported but not officially confirmed, so treat any specific number you see with caution. The takeaway for builders is that "gated" covers a spectrum from a polite delay to a binding order, and the binding end is the one that should shape your fallback plan.
What you can still call, and what it costs
Set the two gated US flagships aside and the field is wider than the headlines suggest. Below is a representative month for an agentic coding workload, 30M input and 6M output tokens, against the models a developer anywhere can sign up for today. We left GPT-5.6 out of the math because it has no public rate card yet. Mythos 5 sits at the top as the reference point you cannot actually reach.
| Model | Input / 1M | Output / 1M | Open weights | 30M+6M month |
|---|---|---|---|---|
claude-mythos-5 (US only) | $10.00 | $50.00 | No | $600.00 |
qwen-3.7-max | $2.50 | $7.50 | No | $120.00 |
glm-5.2 | $1.40 | $4.40 | Yes (MIT) | $68.40 |
kimi-k2.7-code | $0.95 | $4.00 | Yes | $52.50 |
mistral-large-3 | $0.50 | $1.50 | Yes (Apache) | $24.00 |
qwen-3.7-plus | $0.40 | $1.60 | No | $21.60 |
deepseek-v4-pro | $0.435 | $0.87 | Yes | $18.27 |
deepseek-v4-flash | $0.14 | $0.28 | Yes | $5.88 |
A caveat the single column hides: these are not all the same kind of model. Kimi K2.7 Code and DeepSeek V4-Pro are coding specialists, Qwen3.7 Plus is a vision model, and Mistral Large 3 caps output at 256K context against the 1M the DeepSeek and GLM rows carry. Read the table as a price survey of the globally-available field, not a like-for-like swap sheet. One live discount to factor in: Qwen3.7 Max is running a 50% launch promo around $1.25/$3.75, which would roughly halve its $120 row to about $60 while the promo lasts. Two more models worth knowing: MiniMax M3 now lists at $0.60/$2.40 (its seven-day launch promo has lapsed), and ByteDance Doubao Seed 2.1 Pro lists in yuan at roughly $0.88/$4.41, with a Turbo tier at half that. Run your real token mix through the cost calculator before you pick, since the ranking shifts with your input-to-output ratio.
The models a government can't gate
Look at the open-weights column and a quiet point jumps out: the cheapest rows are also the ungated ones. That is not a coincidence so much as a property. An API is a service the vendor runs and can switch off on a government's say-so, which is exactly what happened to Fable 5 and Mythos 5 inside a few hours. Open weights are a file. GLM-5.2 ships its full 753-billion-parameter weights under an MIT license. DeepSeek V4, Kimi K2.7 Code, MiniMax M3, and Mistral Large 3 (Apache 2.0) are all downloadable and self-hostable on any infrastructure in any country.
You cannot recall a file that is already mirrored across HuggingFace, ModelScope, and a few thousand laptops. For a team outside the US that just watched two flagships vanish on a directive, that is the real story in the table. The open-weight option is not only the budget floor, it is the one access guarantee no export order can touch. Renting an open-weight model through a hosted API is still convenient and cheap, but the weights being public means you always have the self-host exit if the API ever gets pulled out from under you.
If you build outside the US right now
Start by separating capability from access risk. If you need a frontier generalist and want to stay on a Western jurisdiction, Mistral Large 3 is the cleanest pick: EU-hosted, Apache 2.0, $0.50/$1.50, and it drops into an OpenAI-compatible client without a rewrite. If raw cost is the constraint, DeepSeek V4-Flash at $0.14/$0.28 is hard to beat and the V4-Pro tier handles the harder reasoning calls for under twenty dollars on our sample month. For coding agents, Kimi K2.7 Code and Qwen3.7 Max are the two that post competitive agentic scores, with GLM-5.2 close behind and MIT-licensed.
Whichever you choose, the lesson the gated-out crowd just learned is worth building in: do not pin production to a single model ID you do not control. A thin routing layer with one hosted frontier model and one open-weight fallback you could self-host costs almost nothing to set up and turns a government directive from an outage into a config change. The American labs may well restore broad access in weeks. Plan as if they might not, and the decision gets easy.
FAQ
Why are GPT-5.6 and Claude Mythos 5 restricted in the US?
A June 2, 2026 executive order created a 'covered frontier model' designation and lets developers give the government up to 30 days of pre-release access. On June 26 OpenAI limited GPT-5.6 to approved partners, and Commerce let Anthropic re-open Mythos 5 only to a small set of trusted US orgs after suspending it June 12.
Are the two restrictions the same kind of rule?
No. The executive order is voluntary on its face. OpenAI calls its GPT-5.6 hold a voluntary delay at the government's request. Anthropic's earlier action was reported as a harder export-control directive covering all foreign nationals. Same headline, different legal instruments.
What can a non-US developer still call?
DeepSeek V4-Flash ($0.14/$0.28) and V4-Pro ($0.435/$0.87), GLM-5.2 ($1.40/$4.40, MIT), Kimi K2.7 Code ($0.95/$4.00), Mistral Large 3 ($0.50/$1.50, Apache 2.0, EU), Qwen3.7 Plus and Max, MiniMax M3, and ByteDance Doubao Seed 2.1 are all outside the US approval process.
Why can't open-weight models be gated?
The weights are already public. GLM-5.2, DeepSeek V4, Kimi K2.7 Code, MiniMax M3, and Mistral Large 3 can be downloaded and self-hosted anywhere. No government can recall a file mirrored on HuggingFace. A vendor's API can be switched off; weights on your own disk cannot.
Is GPT-5.6 priced yet?
Not publicly. OpenAI previewed it to approved partners on June 26 with broad availability promised in coming weeks, but published no rate card. Until it ships there is nothing to budget against, which is its own argument for keeping a non-US fallback wired in.
Sources
- White House - executive order Promoting Advanced AI Innovation and Security (June 2, 2026)
- TechCrunch - OpenAI limits GPT-5.6 rollout after government request
- CNBC - US lets Anthropic re-open Claude Mythos 5 to trusted partners
- Skadden - analysis of the June 2026 AI executive order
- DeepSeek API pricing (V4-Flash, V4-Pro)
- Mistral pricing (Large 3, Apache 2.0 weights)
- GLM-5.2 open weights on HuggingFace (MIT)