Anthropic reports progress on proofs that verify AI model inference
🎧 Listen to this story
Anthropic says it has made progress on cryptographic proofs of inference, which allow a provider to commit to a model and demonstrate that each response was produced by it. Its FlashProve system uses zkSNARKs and, in a test with Llama 4 Maverick, a 400-billion-parameter model, one H100 GPU generated proofs at about 110 tokens per second; each 1,024-token request produced a proof of roughly 19 MB. The company says this required 70 to 95 times the GPU time of running the model on vLLM, while remaining faster than earlier open-source systems.
Anthropic says the proofs hide the model’s weights but not its architecture. The company also released a technical report, protocol description, code with a prover and two verifiers, Lean proofs, and a visualizer, all under the Apache 2.0 license. It describes the work as intermediate progress rather than a completed research agenda and invites others to audit and improve it.
Why it matters · editorial interpretation
The work could provide a verifiable way to link responses to a specific model without revealing its weights, but computational cost and proof size remain practical obstacles. Releasing the code and verifiers allows others to audit whether the approach is sound and viable.
Sources
Research