Tell us when the model changes.

We ask AI providers for pinned versions, public changelogs and advance notice. Add your name.

Sign

Dear OpenAI, Anthropic, Google, xAI, Mistral AI and Meta,

millions of people now build their work on your models. They write code with them, run research on them and put them into products their own customers depend on. They chose a model because of what it could do on the day they chose it.

The model they chose is not always the model they get a few weeks later. Answers get shorter. Instructions get ignored that were followed before. A task that worked yesterday fails today. Sometimes the cause is a bug. Sometimes it is a deliberate trade-off between quality, speed and cost. Sometimes it is just the randomness every language model has.

From the outside, nobody can tell which. That is the problem this letter is about.

What is on the record

We do not rely on rumours. These cases were measured independently or reported by the providers themselves:

  • Anthropic, September 2025. Three infrastructure bugs degraded Claude’s answers between 5 August and 18 September 2025. At the worst point, 16 % of Sonnet 4 requests were affected. Anthropic published a detailed postmortem.
  • Anthropic, April 2026. For about a month, the default reasoning effort in Claude Code was lowered from high to medium to reduce waiting times. It came to light after users reported worse results. Anthropic later called it “the wrong tradeoff”.
  • OpenAI, April 2025. An update made GPT-4o overly flattering. OpenAI rolled it back within days and explained what went wrong.
  • Stanford and UC Berkeley, 2023. Researchers measured clear changes in GPT-4’s behaviour between the March and June 2023 versions of the same model name.

We credit the companies that published these reports. Honest postmortems are the reason this letter can cite facts. But they arrive after the damage, and only when someone decides to write one.

What we ask

  1. Pinned versions. Every model a customer can select maps to a dated version that does not change during its documented lifetime. Names that can change, like “latest”, say so.
  2. A public changelog for how models are served. Changes to quantization, routing, reasoning or context budgets, system prompts and safety layers are logged when they can affect output quality. This applies to apps and subscriptions as much as to the API. If plans get different versions or settings, that is disclosed.
  3. Advance notice. Changes that can lower quality, and retirements of models, are announced before they happen.
  4. Quality you can check. Repeatable quality measurements are published for each served version over time. The reasoning or effort level actually applied to a request is visible to the person who sent it.
  5. Incident reports. When quality drops, you say so, explain why and say when it was fixed, as some of you already do.

This is not a radical demand

EU consumer law already expects something close to it for digital services. A provider may change a digital product only if the contract allows it and customers are clearly informed. If a change impairs the product more than slightly, customers must be told in advance and may cancel. This duty does not apply if the unchanged version stays available at no extra cost (Directive (EU) 2019/770, Article 19). Pinned versions are exactly that.

We are asking you to make this the global default, for every customer and every plan.

This letter accuses no one of fraud. It asks for the transparency that lets people trust what they pay for and plan their work around it.

Signed, the people below.

Sources

Sign the letter

Your e-mail address stays private. A confirmation link makes the signature count.

Only used to confirm your signature. Never published.