Anthropic FlashProve: what it is and how to try it

FlashProve is an open-source research prototype from Anthropic that lets an AI provider prove a specific response came from a specific model, without handing over the model's weights. Anyone can check the proof on an ordinary CPU, but making the proof still takes about 70 to 95 times the GPU time of just running the model, so this is a research release and not something you can switch on in Claude today.

Anthropic announced it on X in a post from @AnthropicAI. The post says FlashProve lets developers "verify whether a given response came from a given AI model." It also names two uses: making a chat harder to hijack, and letting safety evaluators confirm the model they tested is the one customers actually get. A few hours before that, Anthropic engineer @sweis posted the code and the write-up, and pointed out that it was measured on Llama 4 Maverick, a 400B-parameter mixture-of-experts model. The published numbers come from Meta's open model, then, and none of them are for Claude.

The problem it goes after

At the moment you can't check which model answered you. Anthropic's research post, dated November 2, describes a model "secretly substituted in by a routing service for cost reasons," or an attacker tampering with the token stream on its way to you. You pay for the flagship model and trust that you got it.

There is precedent. In September Anthropic alleged that Kimi and DeepSeek users got Claude answers passed off as the apps' own. Model routers and resellers are also turning into a normal layer of the stack, as we covered in Stripe and OpenRouter's AI switching layer. Each extra hop between you and the lab is one more place a cheaper model could be swapped in.

FlashProve uses zero-knowledge proofs. The provider publishes a small commitment to its weights once. After that, each reply can come with a proof that it was computed from those committed weights. That idea comes out of the crypto world, and @mert was quick to post that Anthropic is "now using the same cryptography that zcash was first in the world to pioneer."

What it can do, and what it costs

These figures come from Anthropic's write-up and the GitHub README. They are Anthropic's own measurements.

Anthropic is upfront about the limits. A proof covers an integer version of the model under greedy decoding, so normal sampled replies can drift from it. A 1,024-token context is tiny next to a real chat session. The protocol's soundness is proved in Lean 4, but the fast executable code has no formal guarantees. The post calls the release "intermediate progress, and not a completed research agenda." Lean also did the checking in the Axiom CFSG Lean proof, and it tends to appear whenever someone wants AI work verified by a machine instead of taken on faith.

How to try it

The cheapest way in is the verifier, which checks a proof Anthropic already generated. The README says that route needs Linux, bash and g++ 13 or newer, with under 2 GB of RAM. On a minimal Ubuntu 24.04 machine:

1. As root, install the tools: apt-get update && apt-get install -y --no-install-recommends g++ make python3 unzip
2. Clone the repo, and from here on work as an ordinary user, which is what the README asks for.
3. Build the verifier with verifier/build.sh and wait for the line THE VERIFIER IS BUILT:.
4. Run bin/cp_verify against the files in examples/accepted-proof/, using the full command in the README. The last line should print YES.

If you have an H100 with CUDA 13.2, examples/quickstart-gpu.sh generates real proofs. Running the full-size model takes 280 to 330 GB of RAM, so most people will stop at the verifier.

What it means for a business buying AI

You don't need to install any of this. It's still worth watching, because right now "we use the flagship model" is just a line in a vendor's sales deck, and FlashProve is an early attempt to make that claim checkable. Anthropic built a watermark into Claude's text earlier this year, and this goes after a related question, provenance, from the serving side.

Until proofs like this get cheap, the practical step is plain bookkeeping: know which vendors sit between your team and the model, and get the model name and data terms in writing. If you aren't sure which AI tools your processes already depend on, New Face Design's free process audit maps that out before you add more.