Open source is the exit

On June 12 the Trump administration sent Anthropic a letter. Fable 5 and Mythos 5 went dark for everyone. Six days of outage. No court review, no warning.

On June 26 OpenAI shipped GPT-5.6 — to a handful of government-vetted organizations. Good luck building a business on that. OpenAI’s own blog post said “we don’t believe this kind of government access process should become the long-term default.” They shipped it anyway.

Two of the best-funded AI labs on earth, and in the same month they both proved they cannot ship their own product to their own customers without asking permission. That is not strength. That is a vendor you cannot depend on.

The giants are in trouble

Let’s look at where OpenAI and Anthropic actually stand after June.

They have the muscle today. Fable 5 and GPT-5.6 Sol are strong models. But the moat is gone. The thing that was supposed to justify the valuations — frontier capability nobody else could touch — is now a moving target, and it’s moving toward open source. Z.ai shipped GLM-5.2 the same week Fable 5 went dark, and in real workloads — with proper prompting and memory embedding — it is matching or beating what the closed labs call “frontier.” Not on every benchmark. On the work people actually pay for.

The closed labs are also now squeezed on the demand side by their own government. You cannot ship to a handful of vetted organizations and keep the revenue story. Anthropic took Fable 5 offline entirely — not just foreign users, everyone — because they couldn’t separate access controls in time. That is an engineering organization that cannot ship a feature flag under pressure. OpenAI agreed to gate GPT-5.6 to “trusted partners” and called it a “short-term step.” Read that again: they framed selling to fewer customers as a temporary measure. That is a company that has lost control of its own go-to-market.

And the White House refused a carve-out for the UK. America’s closest ally. An official called the request “completely illogical.” So the policy is not “no China.” It’s no allies either. The US is restraining even its closest partners from getting access to intelligence — and demonstrating, in public, that it will do the same to its own companies whenever it wants.

If you are any enterprise outside the US, the message is clear: American AI is not a reliable supply chain. If you are inside the US, the message is clearer: your vendor answers to a letter, not to you.

GLM-5.2 changed the conversation

Here is what actually shifted in June, and it’s not the ban. It’s that GLM-5.2 showed the world open source can keep up and be competitive.

Is it better than Fable 5 on every benchmark? No. Is it close enough that the difference doesn’t matter for real work? Yes. And here’s the part the benchmark crowd misses: prompt engineering is still on the table. A well-prompted open model with good memory embedding achieves results that match or exceed what the closed labs ship out of the box. The model is half the system. The other half is the craft around it, and that craft works on any provider.

So if the cost of being free and unlocked is one percent on a benchmark, take that trade every time. The benchmark is not the product. The product is what you build with the model, and you can build it on open weights without asking anyone’s permission.

And here’s the part the closed labs can’t answer: GLM-5.2 will become 5.3, then 5.4. The open-source release cadence is relentless — a new version every few months, weights public, community fine-tuning on day one. Fable 5 can’t defend its benchmark lead against that curve. It’s one model, gated, reviewed, and delayed by a government process. Open source is a moving target with a thousand people shooting at it. The gap closes every quarter, and there is no letter that can stop that.

The closed labs know this. That’s why the marketing has shifted from “we’re ahead” to “we’re responsible.” When a vendor starts selling responsibility instead of capability, they’ve lost the capability argument.

When the provider acts like a dictator

Here’s my actual line.

When the provider acts like an authoritarian dictator — deciding who gets access, suffocating the news walls, gating intelligence behind government approval — it’s time to cut off the pipe and own the path entirely.

I’m not interested in funding that. As of today I’m not paying for GPT or Claude subscriptions, and I won’t. Not because the models are bad. Because the models are not mine, and the company controlling them has shown it will pull them on a letter’s notice. I’m not renting infrastructure from a landlord who can be evicted by someone I didn’t vote for.

Open source gives you the path. The weights are on Hugging Face. The licenses are MIT and Apache. You download them, you run them on your own metal, and no Commerce Department letter, no executive order, no “trusted partner” list can take them back. That is not a feature. It is the requirement.

The trade

The closed-AI bet has a failure mode that didn’t exist a month ago: sovereign shutdown. It is not theoretical. It happened. To the best model, from the strongest lab, in the largest market. And the response from the lab was to apologize and ask for a “repeatable process.” That is not a partner. That is a contractor.

Open-source models are not as good on paper. They are good enough in practice, they are yours, and nobody can take them back. With good prompting and the right setup, they are mostly better than what the closed labs produce — because you control the whole system, not just the API call.

It’s time to bring down the giants and empower people. The exit is open. Take it.

A starting point

The weights are the product, and the weights are free. Start at Hugging Face — the licenses are MIT and Apache, and the models run on your own hardware.

DeepSeek V4 Flash is open weights. It is superior to Sonnet 4.6 in every way, and likely to the upcoming Sonnet 5 already, at a fraction of the compute cost. Download it, run it, own it. No subscription, no API key, no letter that can turn it off.

For the interface, Open WebUI is the first-class replacement for Claude Desktop — open source, self-hosted, runs entirely on your hardware. RAG, voice, plugins, the works. Your data does not leave your machine. No telemetry, no “trusted partner” list, no government vetting. You install it, you own it, you control the whole setup. It is what Claude Desktop should have been — minus the landlord.

On mobile, Open MobileUI is the native iOS app that replaces the Claude app. Same principle — open source, private, yours. Your conversations don’t leave your hardware.

I’m not sponsored by any of these projects. I’m giving you the whole post for free because the value is in the path, not in the paywall. Take it and build.

One caveat: most hosted inference providers sit in the States. For now that works. When that pipe is gone — and June showed us pipes get cut — the weights are still public. You can find them on Hugging Face, on self-hosted registries, on mirrors outside US jurisdiction. The model is not the server. The model is the weights, and the weights are already on your machine.