Back to news
launchAnthropic2026-11-04

Anthropic open-sources FlashProve on November 4, cryptographically proving which model produced an output

FlashProve uses zero-knowledge proofs to give mathematical guarantees about specific outputs, letting users verify locally without publishing model weights.

Anthropic is developing an open-source system for verifying which model produced a given AI output. According to a blog post published by the company on November 4, the system, named FlashProve, aims to deliver provable inference: using cryptographic proofs to give mathematical guarantees about specific outputs, showing that they really do come from a designated model. With it, users can confirm whether an answer they received actually came from the model they believe they were using.

FlashProve works roughly like encrypted email signatures, which help recipients confirm a message really came from its claimed sender, except FlashProve resists attempts to disguise one model as another, even when the party doing the disguising is the AI company that developed and runs the model. Researchers note the system could prevent AI companies from quietly swapping in cheaper models to cut costs, and could also stop firms from substituting more expensive models to artificially inflate performance test results. FlashProve can also verify third-party claims that a particular task was completed by an advanced model.

To make this work, AI labs need to provide a fingerprint for each model they publish, which can then be matched against specific outputs and cryptographically verified. Related materials indicate the system has been validated on a 400-billion-parameter mixture-of-experts model, with proof generation throughput around 110 tokens per second on a single H100, proof size around 18.78 megabytes and CPU verification taking roughly 17 seconds, but the GPU time consumed for proof generation is about 70 to 95 times that of ordinary inference, and it currently supports only 1024-token contexts, integer models and greedy decoding, verifying provenance rather than correctness.

FlashProve可证明推理零知识证明Anthropic开源