Anthropic releases FlashProve, a Claude-built system that proves which model produced a response
The zkSNARK prototype proves Llama 4 Maverick outputs at about 110 tokens a second on one H100, roughly 70 to 95 times the GPU time of normal inference, with code and a technical report released.
Why it matters
Users today cannot check that an answer came from the model a provider claims, rather than a cheaper substitute or a tampered token stream. Anthropic says cryptographic proofs of inference fix that while hiding the weights, and could help verify compliance with AI pacing requirements. The system is substantially faster than prior open-source provers, and Claude did much of the optimisation, with the security claims stated in Lean so they can be audited. The open question is whether the overhead falls far enough for routine use.
Key facts
- About 110 tokens per second on a single H100 for Llama 4 Maverick (400B MoE)
- Proof size about 19 MB per 1,024-token request
- About 70-95x the GPU time of running the model on vLLM
- Security claims stated in Lean
- Five artifacts released, including code and a technical report