Breaking01 / AI
Anthropic open-sources FlashProve to verify AI model outputs
The system checks which model produced a response while keeping the weights hidden, but Anthropic has not given a schedule for using it in Claude services.

Anthropic released FlashProve, an open-source system for proving which AI model produced a response, on 2 Nov 2026, with a company-published speed of about 110 tokens per second on one H100 GPU for Llama 4 Maverick.
Anthropic said it released FlashProve's code, documentation and proofs in Lean, a formal verification language, under the Apache 2.0 licence. The system lets a user check whether a response came from the weights the provider committed to, without access to those weights, the learned parameters of a model. Its core technique is the zkSNARK, a succinct zero-knowledge proof that shows a computation ran correctly without revealing anything else.
Measurement conditions. According to company-published figures, the test model was a 4-bit quantised version of the 400B-parameter Llama 4 Maverick. The hardware was one H100 GPU, and the proof for a 1,024-token request was about 19 MB.
The GPU time for proving was about 70 to 95 times that of processing a prompt of the same length with vLLM, a widely used open-source library, which is also a company-published figure. Anthropic said earlier open-source systems were orders of magnitude slower.
Anthropic wrote in its research post that the result is intermediate progress and "not a completed research agenda". Claude, the company's AI model, produced the improvements, and the security guarantees are open to audit through Lean theorems, according to the post.
In its Frontier Safety Roadmap, Anthropic set a goal of building a provable inference prototype by 30 Sep 2026. The company said it met that goal through a separate project that uses signing and recomputation on a trusted server. FlashProve is research that pursues the same goal by a different method.
Anthropic has not disclosed a schedule for applying FlashProve to Claude services.