News · 2026-07-25

The open-weights letter doubled to 50 signatories - Google and OpenAI signed, Anthropic did not

The industry letter asking Washington not to restrict open-weight AI models doubled its signatory list from 25 organizations to 50 within roughly a day of launch, and the new names include Google and OpenAI. Anthropic does not appear on the list. The letter, hosted on Microsoft's corporate-responsibility site and dated July 24, 2026, asks policymakers to expand compute access and shared training resources while avoiding what it calls premature restrictions on models that anyone can download and run.

Key facts

• The signatory count went from 25 to 50 between the launch PDF and the live page, both dated July 24, 2026.
• Added names include Google, OpenAI, AMD, Cisco, Cloudflare, Cohere, GitHub, Ollama, Nous Research, Prime Intellect, Nebius, Baseten, Fireworks AI and Palo Alto Networks.
• Anthropic is absent. So are xAI, Amazon, and Apple.
• Primary source: the coalition letter hosted by Microsoft, compared against the original launch PDF hosted by NVIDIA.

When we covered the letter yesterday, the roster was still filling in. It has now settled at fifty organizations spanning chipmakers, clouds, model labs, open-source foundations, payment companies and security vendors: AI21, AMD, Andreessen Horowitz, Arcee AI, Black Forest Labs, Block, Box, Cloudflare, Cohere, CrowdStrike, Dell, DoorDash, GitHub, Google, Hugging Face, IBM, the Linux Foundation, Meta, Microsoft, Mistral, Mozilla, Nous Research, NVIDIA, Ollama, OpenAI, Palantir, Perplexity, Prime Intellect, Replit, ServiceNow and Y Combinator, among others.

The absence that got attention is Anthropic's. That absence is real, and it is worth understanding accurately rather than dramatically.

What the letter actually argues

The letter is narrower than "open source good." It defines open weight models as models that anyone can, in its words, "download, inspect, modify, and run on their own infrastructure" - a definition worth reading closely, because it is doing policy work. Open weights, in this framing, are not a licence or a philosophy. They are a distribution fact: the numbers that make up the model are in your hands, on your hardware. Our explainer on what open weights really mean walks through why that distinction matters more than the word "open" on a model card.

From there the letter argues for what amounts to defensive pluralism. If more independent parties can audit, red-team, adapt and self-host capable models, the argument goes, the ecosystem is more resilient than one that routes everything through a handful of closed providers - each of which becomes a concentrated failure point. The policy ask is expanded compute access, shared training assets, and restraint on sweeping restrictions.

Notably, the letter does not pretend the tradeoff away. It states plainly that once weights are released they are beyond the original developer's control, and that modified versions are hard to trace or reverse. That is the strongest argument against its own position, and the coalition puts it in the text.

The distillation clause is the part that let OpenAI sign

The most consequential paragraph is about distillation - the practice of training a smaller model on a larger one's outputs. The letter distinguishes ordinary output-based distillation, which it says has legitimate uses in improvement, evaluation and validation, from unlawful extraction from closed systems, which it says should face targeted legal and commercial remedies rather than blanket restrictions on open models.

That clause is why a company like OpenAI can sign without contradicting itself. Signing this letter does not commit OpenAI to open its frontier weights, and it does not require the company to drop its objections to account-fraud-driven extraction of its models. It is a statement about what governments should regulate, not about what any signatory will publish. The distinction matters as Washington weighs sanctions tied to industrial-scale distillation.

Anthropic's position, steelmanned

Anthropic has published no direct response to the letter. What its own material does show is a conditional position rather than a blanket one. In its writing on third-party testing, Anthropic argues that most current AI systems are likely safe to disseminate openly, and credits openness with helping security research. Its concern is the future frontier: once a model can meaningfully assist serious misuse, the controls Anthropic relies on - monitoring, classifiers, customer vetting, rapid updates - do not travel with a downloaded weights file. The company also concedes that, as a proprietary-model business, it lacks the standing to decide alone what may be released, and proposes independent testing and narrowly defined thresholds instead.

Its ASL-3 protections post makes the mechanism concrete: Anthropic says it runs more than a hundred security controls specifically to keep weights from being stolen, because copied weights would let someone run the model without any of its deployment safeguards.

So the real dispute is not open versus closed. The coalition treats openness itself as a security control - more eyes, more auditors, less concentration. Anthropic treats controllable deployment as indispensable once capability crosses a threshold. Both can be arguing in good faith and still be irreconcilable.

The honest caveat

The live page does not display a revision timestamp, so while the roster change from 25 to 50 is verifiable from the two primary artifacts, the exact hour each name was added is not. Neither Google nor OpenAI published a standalone announcement of signing; the confirmation is their appearance on the letter itself. And the popular framing of "every major lab against Anthropic" does not survive contact with the list - xAI, Amazon and Apple are missing too.

Primary source, verified: read the paper →

Ground Truth. (2026, July 25). The open-weights letter doubled to 50 signatories - Google and OpenAI signed, Anthropic did not. Ground Truth. https://groundtruth.day/news/open-weights-letter-doubles-to-50-anthropic-absent.html

@misc{groundtruth:open-weights-letter-doubles-to-50-anthropic-absent,
  title  = {The open-weights letter doubled to 50 signatories - Google and OpenAI signed, Anthropic did not},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/open-weights-letter-doubles-to-50-anthropic-absent.html}
}

Topics: open-weights · policy · industry · anthropic · openai · google