Anthropic releases FlashProve for verifiable model inference

Anthropic’s Frontier Red Team has released FlashProve, an open-source research prototype using zkSNARKs to prove that a response was generated by a specific committed model. The proof does not disclose the model’s weights. The system is designed to address a verification gap in API services, where users currently cannot independently confirm which model produced an answer or whether token output was altered.

On Llama 4 Maverick, a 400B mixture-of-experts model with about 17B active parameters per token, FlashProve reaches about 110 tokens per second using one H100 80 GB GPU and 24 CPU cores. Proofs are 18.78 MB and take about 17 seconds to verify on 24 CPU threads. The prover requires roughly 290 GB of host RAM, while verification does not require a GPU. A request is limited to 1,024 tokens.

The provider first publishes a cryptographic commitment to one set of weights, a process taking about 39 minutes and covering approximately 16 million requests. Proofs are then generated for individual responses. Anthropic says the protocol has post-quantum security properties and formalized completeness, soundness and zero knowledge in Lean. The trade-off is substantial cost: proving requires about 70 to 95 times the GPU time of ordinary vLLM inference.