Anthropic's FlashProve proves which AI model replied

1. What was announced

On November 2, 2026, Anthropic published FlashProve, research on using cryptography to prove that an AI response came from a specified model. Its Frontier Red Team worked on it with researchers from EPFL, MIT and others. The problem it addresses is that a user cannot tell from the output alone which model actually replied, so a routing service that quietly swaps in another model to cut costs, or an attacker who tampers with the token stream, goes unnoticed. Anthropic says this is intermediate progress, not a completed research agenda (Anthropic announcement).

The mechanism is that the provider first publishes a short fingerprint (a commitment) of the model's weights, then attaches to each response a proof file showing that the model with those weights produced this output from this input. The recipient does not need the weights and can check the proof in seconds on a CPU with no GPU; a response from a different model fails the check. The proof reveals nothing about the weights, though the model's architecture is not hidden. The code and the formal proofs in Lean are released under Apache 2.0 (Anthropic announcement, GitHub).

In Japanese media, GIGAZINE covered it on November 4 as a countermeasure to the current situation in which one cannot rule out a service quietly using a lower-performance model (GIGAZINE). ITmedia NEWS listed uses such as letting a third-party evaluator confirm that the model it tested is the one actually being served (ITmedia NEWS).

2. Why it is not ready for use yet

Anthropic itself lists the limits. The test used a 4-bit version of Meta's Llama 4 Maverick (a 400B-parameter mixture of experts); one H100 produces proofs at about 110 tokens per second, which is roughly 70 to 95 times the GPU time of running the model normally. The context is limited to 1,024 tokens, much shorter than real chats, and the proof covers the output of an integer version of the model under greedy decoding (always choosing the most likely token). The GitHub README also states that this is a research artifact that will not be maintained (Anthropic announcement, GitHub). ITmedia NEWS notes that because the proof is generated after the response, it detects tampering after the fact rather than preventing it (ITmedia NEWS).

Our view For small businesses, the near-term meaning is less the technology itself than the idea of verifying what is inside an AI service becoming standard. Companies that stack cheap AI services will end up using them for customer inquiries and quotes without being able to say which model is really running, and will be asked to explain only after something goes wrong. What can be checked today is whether the vendor documents which model it uses and under what conditions it switches. If a vendor cannot show that, we would treat its use in important work as lower priority.

3. What small businesses should check now

List the AI services used internally and the AI used for customer replies and quotes
Check whether each vendor names the model it uses in the contract or specification
Ask each vendor whether you are notified if the model is switched mid-contract
For AI used in important decisions, set a step in which a person reviews the result
This research is at prototype stage, so do not rush adoption; review vendors' responses every six months

Our outlook

Over the next six months to two years, we expect AI that searches for and compares shops and suppliers on behalf of customers to spread, and later for AI agents to negotiate terms directly with each other. At that point, a way to check that the other AI really is the model or business it claims to be becomes a building block of trust in deals. FlashProve is one entrance to that, and speed and context length still stand between it and practical use. Even so, companies that can explain what their AI uses and how it answers will be easier for both people and AI to choose. Now is the time to lay the groundwork: make product and price information machine-readable and keep records of how your AI is used.

Questions this article answers

What is FlashProve, and does it affect how we use AI?

It is a research prototype Anthropic released on November 2, 2026 that uses cryptography to prove an AI response came from a specified model. It is an unmaintained research artifact, not a product you can use now. For now, it is enough to ask your AI vendors which model they use and under what conditions they switch it.

How can we check that an AI service is not secretly using a cheaper model?

Anthropic itself says that output quality alone is hard to judge by. What you can do today is check whether the contract or specification names the model and promises notice on a switch. Until proofs like FlashProve become practical, also keep a step in which a person reviews results for important work.

Sources

Related Q&A

Related services

If you want to try this in your own business

To see how the topics in this article would apply to your own operations, and what it would cost, send us a message through the contact form.