AI Tech Flash ── 2026-11-04
Anthropic prototypes proof of which AI model answered
Anthropic releases FlashProve, a cryptographic proof of which model produced a response.
What changes
Users may gain a way to check that an AI response came from the model that claimed to produce it. According to Anthropic, the proof does not reveal the model's weights, the internal numbers that define its behavior.
However, Anthropic calls the post intermediate progress, not a completed research agenda. According to Mezha.net, there are no plans to deploy the system at Anthropic or any other lab.
Why it matters
According to Anthropic, users today have almost no way to confirm which model produced a response. A routing service could swap in another model to cut costs. An attacker could tamper with the token stream, the sequence of text units a model outputs.
Anthropic also cites independent evaluators who need to know the model they test is the one being served. It says providers could also use proofs to show they meet commitments and legal obligations.
What it is
FlashProve is an open-source prototype that Anthropic calls a system for provable inference. A provider commits to a model and then proves that each response was produced by that model.
It uses a zkSNARK, a cryptographic proof that is short, quick to check and reveals nothing beyond the claim. According to Anthropic, the proofs hide the model's weights, but hiding the model's architecture is not a goal.
Who and where
The authors include Roy Rinberg and Alex Allain of Anthropic, with researchers from EPFL, MIT and other institutions. Anthropic released the code, Lean proofs and documentation under the Apache 2.0 license.
It released five artifacts: a blog post, a technical report, a protocol specification, executable code and Lean proofs. According to Mezha.net, parts of the code are on GitHub. Mezha.net also reports there are no deployment plans.
Why it was built
According to Anthropic, most frontier models keep their weights private. Even with the weights, checking a response would mean running the model, which needs more computing than a typical user or evaluator can afford.
Anthropic therefore aims to let a user with limited computing and no knowledge of the weights verify that a response came from the requested model. It says this also bears on pacing AI development and on AI regulation.
How it works
According to Anthropic, the provider sends a short proof with each response, a few megabytes, ideally checked in seconds on a CPU. If the response deviates from the specified model, the check fails.
In Anthropic's announcement, the test used Llama 4 Maverick, a 400B-parameter mixture-of-experts model, in 4-bit quantization, which lowers numeric precision. One H100 GPU generated proofs at about 110 tokens per second, at about 19 MB per 1,024-token request. That is about 70 to 95 times the GPU time vLLM needs for prompts of the same length.
Anthropic says Claude's optimization choices are hard to evaluate, so it provides top-level security statements in Lean, a formal verification language.
Key points
- Anthropic released FlashProve, a prototype that proves which model produced a response.
- In Anthropic's announcement, one H100 reaches about 110 tokens per second with about 19 MB proofs.
- Anthropic calls it intermediate progress, and Mezha.net reports no deployment plans.
Conclusion
FlashProve is a prototype for verifying which model produced a response while keeping weights hidden. The performance figures come from Anthropic's announcement. Anthropic describes it as intermediate progress, and Mezha.net reports no deployment plans.
Quotes and sources
“Cryptographic proofs of inference solve this problem: they let a model provider commit to a model and then prove that each response was produced by that model.”── Anthropic, 2026-11-03
“a single H100 GPU generates proofs at around 110 tokens/second, with a proof size of ~19 MB per 1,024-token request”── Anthropic, 2026-11-03
“This research post is meant to share intermediate progress, and not a completed research agenda.”── Anthropic, 2026-11-03
“Research is ongoing, and there are no plans to deploy the system at Anthropic or any other lab.”── Mezha.net, 2026-11-04
Source articles
Written by AI (Claude) from the cited articles, then checked by Jev (TypeSafe's judgement model), a Gemini fact-check and a verbatim check of every quote.