Anthropic tests a way to prove you got the AI model you paid for
Anthropic has released FlashProve, a free research prototype that lets an AI provider prove a reply came from the model it promised, without revealing that model's secrets. It is early, but it points to a check buyers have never had.
The short version
- Anthropic released FlashProve, a free research prototype that attaches a proof to each AI reply showing which model produced it.
- Anyone can check the proof on an ordinary computer in seconds, and the model's inner workings stay secret.
- It is not a product yet, but it is the first serious step toward checking that you get the AI you pay for.
The problem it tackles
When your company pays for an AI service, you usually trust the label. The provider says the answer came from its top model. You can't see inside its servers to confirm it.
That matters more as AI prices spread out. A top model can cost many times more to run than a small one. A provider, or a reseller in the middle, could quietly send some requests to a cheaper model. A reply would look much the same, and you would have no way to tell.
Regulators face the same gap. If a model passed a safety review, nobody outside can confirm that the same model is answering customers today.
What FlashProve does
FlashProve, published by Anthropic on GitHub, works in three steps.
First, the provider publishes a commitment ID. Think of it as a sealed fingerprint of one exact model. The model itself stays hidden.
Second, for each reply, the provider's computer produces a mathematical proof. The proof shows the reply came from the model behind that fingerprint.
Third, anyone can check the proof on an ordinary computer in a few seconds. They don't need the model or any special hardware.
Anthropic has also used a computer proof checker, called Lean, to confirm the method's math. A provider that tries to pass off another model's output gets caught, except in cases so rare they can be ignored.
The limits today
This is a research project, and Anthropic says so plainly. The code is marked as a research artifact. It is not maintained and won't take outside contributions.
It runs on one model so far: Llama 4 Maverick, a large free-to-download model. It handles only short requests, about a page or two of text. Making proofs needs one of NVIDIA's most expensive AI chips, set aside for the job.
There is a practical catch too. The fingerprint is only useful if you get it from someone you trust. Anthropic's own guide says never to take it from the server being checked.
The code is free to use under the Apache 2.0 license. Other providers can study it or build on it.
What it means for companies in Indonesia
Nothing changes in your contracts this month. But the direction matters for anyone buying AI, especially through resellers and local partners.
Indonesian banks and insurers increasingly buy AI through a chain of suppliers. Each link is a place where the model could change without notice. Proofs like these could one day let you check, reply by reply, that the model you approved is the one serving customers.
It also fits where regulation is heading. Rules that require a model to pass a review only work if you can show the reviewed model is the one in use.
What to do next
- Write the exact model name and version into your AI contracts, with a duty to tell you before any change.
- Ask suppliers and resellers how they prove which model answers your requests today.
- Keep a small set of test questions and run them each month, to spot sudden changes in quality.
- Ask your main AI provider whether it plans to offer proofs of which model replied.
- Have your IT security team follow FlashProve and similar work, and revisit the topic in a year.