The US Government Now Decides Who Gets to Use Frontier AI. Open Source Is the Escape Hatch.
GPT-5.6 Sol and Claude Mythos 5 both shipped this week — to government-approved lists only. The Commerce Department is building a new regulatory regime on the fly, with no legislation and no public debate. If you build with AI, open-source models just became existential.
Something happened this week that didn't get enough attention.
On Friday, the US government lifted its block on Anthropic's most powerful model — Claude Mythos 5. Two weeks earlier, Commerce Secretary Howard Lutnick had imposed export controls on it, citing concerns that the model could be "jailbroken" for malicious purposes. The shutdown happened. It was real.
Now, the block is partially lifted. Mythos 5 goes to "more than 100 US institutions" — companies and agencies on a government-approved list. The weaker sibling, Fable 5? Still locked. No word on when consumers get it back.
The same day, OpenAI released GPT-5.6 Sol — their new flagship. Also to "a small group of trusted partners whose participation has been shared with the government."
Read that again. The government now decides who gets to use frontier AI models.
Not the companies. Not you. A regulatory framework that's being built on the fly, in letters between Commerce Department officials and AI executives, with no legislation, no public debate, and no clear endpoint.
If you build software, automate workflows, or just use AI to get work done, this matters more than any benchmark score.
What Actually Happened
Let's be precise, because the details matter.
Anthropic's situation: Two weeks ago, the Trump administration imposed export controls on Claude Mythos 5 and Fable 5. The models went dark. Amazon and other companies had warned they could be jailbroken. On Friday, Lutnick sent a letter to Anthropic's chief compute officer Tom Brown, writing: "I have determined that appropriate safeguards are in place to permit certain trusted partners to access the Claude Mythos 5 Model."
The key phrase: "trusted partners." A license is no longer required to transfer Mythos 5 to "entities identified in Annex A" — a list you and I will never see. Anthropic committed to "work with the U.S. government on protocols and standards and releases" for its models going forward.
OpenAI's situation: GPT-5.6 is a family of three models — Sol (flagship), Terra (mid-tier), and Luna (fastest/cheapest). From the system card: "As part of our ongoing engagement with the U.S. government, we previewed our plans and the models' capabilities ahead of today's launch. At their request, we are starting with a limited preview for a small group of trusted partners whose participation has been shared with the government, before releasing more broadly."
"At their request." The government asked OpenAI to restrict access, and OpenAI complied.
This is a new regulatory regime. It's not Congress. It's not the EU AI Act. It's executive-branch export controls, applied to software, with a vetted access list maintained by the Commerce Department.
The Technical Side (Which Is Actually Good News)
Buried under the policy noise, GPT-5.6 Sol is a legitimate step forward.
The system card classifies all three models as High capability in both Cybersecurity and Biological/Chemical risk — but not Critical, their highest threshold. Sol and Terra can find vulnerabilities and pieces of exploits, but couldn't carry out autonomous end-to-end attacks against hardened targets in testing. They're better defenders than attackers, which is at least the right direction.
OpenAI threw 700,000 A100e GPU hours at automated jailbreak finding. They added activation classifiers that watch the model in real-time during generation and can intervene to stop unsafe outputs mid-sentence. The safety stack is genuinely more sophisticated than anything they've shipped before.
And the capability bump is real. The most interesting number came in the second-to-last paragraph of the announcement: GPT-5.6 Sol will run on Cerebras at 750 tokens per second starting in July.
For context, Claude Opus 4.8 runs at about 55 tokens/sec. Fast mode gets you ~102. 750 tokens/sec on a frontier model is absurd — fast enough that an AI agent reading a codebase, running tests, and iterating feels fundamentally different. Speed compounds with agentic workflows because every step that used to take 10 seconds now takes 1.5.
So yes, the models are getting better. That's not the problem.
The Problem Is Who Holds the Keys
Here's the uncomfortable part.
When OpenAI or Anthropic decides who gets their models, that's a business decision. You might not like ChatGPT's pricing or Claude's rate limits, but you can negotiate, switch providers, or run something else.
When the Commerce Department decides, it's something else entirely. There's no negotiation. There's no alternative provider within US jurisdiction. There's a list — Annex A — and you're either on it or you're not.
The Hacker News thread on this hit 996 comments in 11 hours. The top comment cuts to the chase:
"This is regulatory capture in action. This will make it hard/impossible for new vendors to come into the market and only established companies will get to play. What does this mean for open source? Will it become illegal to download weights? What about train your own?"
These aren't paranoid questions. The precedent is being set right now. If Mythos 5 — a software model — can be subject to export controls, what stops the next framework from applying the same logic to open-weight models that match frontier capability? GLM-5.2 already ties GPT-5.5 on agentic benchmarks. It's MIT-licensed. What happens when "open weights that are too good" becomes a policy category?
Europe isn't helping. The EU signed onto what commenters are calling "Pax Silica" — voluntarily ceding the frontier model space to US incumbents while banning Chinese models. So European developers get to rent whatever Washington allows them to rent, from whichever companies the Commerce Department approves. They're choosing dependency at the exact moment they're trying to reduce it in every other domain.
One HN commenter put it bluntly: "At the same historical turning point when Europeans are finally waking up to the need to become less dependent on the US for weapons and security, they are immediately choosing to become dependent on the next layer of critical infrastructure."
That's not wrong.
Why Open Source Just Became Existential
Here's where this connects to anyone actually building things with AI.
The entire appeal of open-weight models — GLM, DeepSeek, Qwen, Kimi, MiniMax — was always more than "they're free." They're sovereignty. You download the weights. You run them on your hardware. No API. No rate limits. No government access list. No company can shut them off, and no Commerce Department letter can revoke them.
That used to be a nice-to-have. An ideological preference. A hedge.
This week it became the rational default.
Think about what a government access list actually means in practice. If your entire product depends on GPT-5.6 Sol, and your company isn't on Annex A, you're done. Not "slower" — done. You're waiting for approval from a process that's being improvised. Meanwhile, a competitor who built on GLM-5.2 or DeepSeek V4 — models that are open-weight, MIT-licensed, and running on their own GPUs — doesn't even notice.
The gap between "permitted access to frontier models" and "models you control entirely" just widened into a strategic moat. And it's the open-source side that has the moat.
The counterargument is compute. Running frontier-scale open models takes serious hardware. A 744B parameter model like GLM-5.2 needs 8x H200s minimum. That's $200K+ in GPUs. Not everyone has that.
But the trend lines are moving fast. VibeThinker-3B just scored 94.3 on AIME26 — competitive with 671B models — using 3 billion parameters. Distillation and efficient architectures are collapsing the hardware requirements for capable models. The question isn't whether you'll be able to run frontier-competitive models locally. It's when.
And on the policy side, the open-source community has one advantage that closed labs don't: you can't put export controls on something that's already been downloaded. Once GLM-5.2 is on Hugging Face, it's everywhere. The weights are on laptops in 40 countries. The horse has left every barn simultaneously.
The Real Stakes
Let me be clear about something: the government has legitimate concerns. Frontier models that can find zero-days, troubleshoot virology protocols, and accelerate biological research are genuinely dual-use. The Cybersecurity capability assessments in GPT-5.6's system card aren't theater — Sol is meaningfully better at offensive security than GPT-5.5 was. That deserves attention.
But the way this is happening should bother everyone.
There's no legislation. No public comment period. No congressional debate about where the line is between "dual-use" and "general purpose." A Commerce Secretary is writing letters to tech executives, and those letters are becoming de facto law. The framework is being built "on the fly," as Semafor put it. And the people being locked out — developers, startups, entire countries — have no recourse.
The result is a two-tier system. Companies on the approved list get the best models. Everyone else gets whatever's left over, whenever it's deemed safe enough, from whoever's willing to sell it to them.
Except — and this is the part that matters — for open source. Open weights don't ask permission. They don't need Annex A. They just work, for everyone, everywhere, at whatever speed your hardware allows.
If you're building AI products right now, ask yourself a simple question: do you want your infrastructure to depend on a list you can't see, maintained by people you didn't elect, through a process with no clear rules?
Or do you want to download the weights and move on with your life?
The answer seems obvious. But it only stays obvious as long as open-source models remain competitive — which means it's never been more important to support, use, and build on them.
If this resonated, CopperRiver is a desktop AI assistant for Mac that runs on open-source models — GLM, DeepSeek, Qwen, Kimi, MiniMax. No government access list required. It browses the web, runs terminal commands, reads your files, and automates tasks. Plans start at $9/mo.