
Anthropic releases FlashProve, cryptographic proof that an AI answer really came from that model
Anthropic has released FlashProve, an open-source system that uses zkSNARKs to prove an AI response came from the promised model without revealing its weights.
When you chat with ChatGPT or Claude, you accept the answer with no way to confirm it really came from the model the provider claims. Anthropic wants to change that. On November 2, 2026, the company released FlashProve, an open-source prototype that uses cryptography to prove each response was generated by the promised model, without revealing the weights that stay secret.
1. The problem it addresses
Anthropic's blog post puts it plainly: a user today cannot check whether responses came from the specific model a developer claims, or from a cheaper model secretly substituted by a routing service to cut costs, or even from an attacker tampering with the token stream. Judging the output on quality alone is not reliable.
FlashProve uses a type of cryptographic proof called a zkSNARK, the same family of proofs behind blockchains like Zcash and Ethereum L2 networks. The model provider first publishes a commitment, a cryptographic fingerprint that pins down one exact set of weights without revealing them. After that, every response can carry a proof that it was the committed weights that produced it.
2. How it works
The system has two phases. In the offline phase, the provider publishes the weight commitment, about 6.5 MB and valid for roughly 16 million requests, according to the technical details in the blog post. In the online phase, for each request, the provider generates a proof file stating that the committed model, run on this prompt, produced exactly these output tokens.
Anyone holding the commitment, the prompt and the reply can check that proof without the model weights and with limited compute. Anthropic notes in its post that the proofs hide the model's weights, but not its architecture, and that hiding the architecture is not a goal of this system.
3. Performance in testing
Anthropic tested FlashProve on Meta's Llama 4 Maverick, a 400-billion-parameter mixture-of-experts model, in 4-bit quantized weights. According to the Anthropic blog post, a single Nvidia H100 GPU generates proofs at around 110 tokens per second, with a proof of roughly 19 MB per 1,024-token request. Anthropic itself says the process takes about 70 to 95 times the GPU time of running the model normally through vLLM.
The front-page figures in other reports differ slightly: Silicon Review, citing Anthropic's announcement, records 72 to 94 times the vLLM GPU time and an 18.78 MB proof. Verifying a proof is fast: about 17 seconds on an ordinary CPU across 24 threads, no GPU and no weights needed, per both sources. Anthropic says earlier open-source systems for the same task are orders of magnitude slower, and that the first proof of this same model once took 14.3 hours for 1,000 tokens, per The Next Web.
4. Claude helped, and Lean proofs keep it honest
Notably, Anthropic says Claude itself helped build the system, including finding novel ways to apply state-of-the-art tools from the cryptography literature. Because Claude's choices can be hard to audit in a large complex system, the team wrote the key security guarantees as machine-checked formal proofs in Lean. The code, the protocol specification and the Lean proofs are released under the Apache 2.0 license.
5. The uses Anthropic sees
The Next Web lists several practical uses. Independent evaluators who test labs' models could confirm they are actually testing the model being served. Customers of routing services could verify they got the model they paid for, not a cheaper or distilled one, an accusation labs have traded this year. Regulators could check that a provider's chips are serving approved models, which is why Anthropic ties the work to Dario Amodei's call to pace the frontier of AI.
6. The limits you should know
This is a research prototype, not a product. It handles contexts of only 1,024 tokens, far below what chatbots use. It works only for mixture-of-experts models. It covers greedy decoding only, not sampled output. It does not protect against timing attacks either. The Next Web reports, based on TechCrunch, that there are no plans to deploy it at Anthropic or any other lab. The GitHub repository itself is labeled a research artifact that is not maintained.
What it means for you
FlashProve changes nothing about the products you use today, but it points to a direction: one day you may be able to verify which model actually answered you.
- Right now, there is no model guarantee. Routing services can swap models silently. If you pay for a specific model, you rely entirely on the provider's reputation.
- Watch the release notes. When a provider swaps the model behind the scenes, your work quality can shift with it. Release notes are the only signal most users get.
- For companies buying AI, ask about verifiability. If you buy APIs or routing for sensitive work, cryptographic proofs like this become a reasonable selection criterion in the future.
- Do not rush to use FlashProve. It is a prototype, capped at 1,024 tokens, and unmaintained. For developers, its value is as an open-source reference implementation plus the Lean proofs.