Their position on open-weights models sucks

I assume by now we've at least heard about the latest blog post from Dario of Anthropic. He claims as he has for years that open-weights models are a possible way for the Chinese and other governments to use our hard earned models for possible repression and geopolitical/military terror.

"We worked real hard distilling and stealing all of humanity's information, it's ours, you can't have it! *spits in disgust*"

That's a direct quote from Dario by the way.

I'll be going over some of the practices and problems that I've seen, and will try to use as much data and analysis as possible from this point forward. Buckle up. If anyone from OpenAI is reading, you're in for a treat.

Nobody signed it for them

On July 24 a letter went out. "Open Weights and American AI Leadership." The pitch is not complicated. Downloadable model weights are strategic infrastructure for this country. They're how you get competition, security research, and an industry that isn't only three companies.

Twenty-five names on it day one. Microsoft. NVIDIA. Meta. IBM. Hugging Face. Mistral. Mozilla. CrowdStrike. Palantir. Perplexity. Y Combinator. Dell. Andreessen Horowitz.

A day later it was fifty. AMD. Cisco. Cloudflare. GitHub. Block. Ollama. Google.
And OpenAI.

For my friends at OpenAI

You've been the punchline for three years. The "Open" in the name turned into a joke. I've made that joke.

You signed it.

Anthropic didn't. Not in the first twenty-five, not in the fifty. The safety company, the one whose entire brand is being the only adults in the room, looked at a letter saying open models are good for America and decided to sit that one out.

The receipt

Meta shipped Llama. Google shipped Gemma. xAI released Grok's weights. Mistral's entire existence is open weights. Microsoft has Phi. NVIDIA has Nemotron. OpenAI has GPT-2, Whisper and gpt-oss.
Anthropic has released nothing. Not one open-weights model. Not a small one, not an old one, not a research checkpoint from 2022 that couldn't hurt a fly if you strapped it to a missile.

Every single competitor has done it at least once.

The caveat

And it's always caveats. Never a straight no, always a caveat.

He'll tell you he has "never advocated for a ban on open-weights models." Technically true. He'll tell you open models are "a public good" when they don't carry dangerous capabilities. He said that. Out loud, last week.

Cool. You've been running a frontier lab for five years. You are sitting on a filing cabinet of models that are nowhere near the frontier anymore and could not be dangerous if you tried. Where are they?
The caveat is always sized exactly right. Open models are a public good, except the capable ones, and Anthropic only makes capable ones, so the public good stays theoretical.

And notice the timing on all of it. The caveats didn't show up when they were behind OpenAI, they showed up when they got in front.

Six weeks ago this stopped being a hypothetical

June 12, 2026. The US government told Anthropic to cut off Claude Fable 5 and Mythos 5 for all foreign nationals, after somebody jailbroke the guardrails off. Anthropic pulled both models worldwide.

Then watch the week that follows, North Mini Code shipped three days before the order. Kimi K2.7-Code shipped the day of. Zhipu opened GLM-5.2 the next day under an MIT license, do whatever you want with it, forever. Inside a week GLM-5.2 was on top of the open leaderboards, running on Huawei silicon, and it was the best model a very large chunk of this planet could legally touch.

Those couldn't be recalled.

That's the part nobody at Anthropic wants you to sit with for too long. A closed model isn't yours. It's a rental, and the landlord isn't even the company you're paying every month. Your access to Claude lasts exactly as long as no government decides otherwise.

Their answer is that the shutdown was the system working correctly. And it was. That's the whole problem. The kill switch worked perfectly.

And it isn't cheap either

Opus 5 runs $25 per million output tokens. DeepSeek V4 Pro is 28 cents. That's 89 times. GLM-5.2 is $4.40.

You can make the capability argument at me, they aren't the same models, and cost per finished task closes some of that gap. The Chinese labs are also probably having losses on domestic GPUs to buy market share.

Fine. It's still 89 times. And the single thing holding that number where it is, is that you cannot download the alternative from the company charging you the 89. They would very much like to keep it that way, and they've written several thousand words explaining why that's for your own good.

What this means

Anthropic is the only frontier lab that has never released an open-weight model. It's the only one whose name is missing from a letter saying open weights are good for the country. And what it offers instead is a rule that open models are fine when they're weak and dangerous when they're strong, which means "never" while managing to sound like "sometimes."

That's what actually gets me. It cost nothing. Write your name, ship nothing, change nothing, go back to work. Half the industry did it, including the company that spent three years getting mocked into the ground over the word "Open" sitting in its own name.

Anthropic couldn't even do that.