Anthropic Open-Sources FlashProve Prototype to Cryptographically Verify AI Inference Origins via zkSNARKs

Anthropic's Frontier Red Team announced FlashProve on its official research blog on November 2. FlashProve is an open-source research prototype designed to generate cryptographic proofs using zkSNARKs. This mechanism allows users to verify that output tokens were generated by a specific, pre-committed model without requiring providers to expose proprietary weights, addressing scenarios where routing services might secretly swap models or attackers modify token streams. In benchmarks using a 4-bit quantized version of Meta's Llama 4 Maverick MoE model, FlashProve's proof workload scales only with the approximately 17 billion active parameters per token. Running on a single NVIDIA H100 with 24 CPU cores, the prover achieves a throughput of approximately 110 tokens per second. Verification requires downloading about 6.5MB of commitment data and runs on standard CPUs within tens of seconds. However, proof generation requires approximately 70 to 95 times the GPU compute time of vLLM inference. The system protocol includes optimizations devised by Claude and formal verification of top-level properties via Lean. Current constraints include a context window limit of 1,024 tokens, support restricted to quantized models with greedy decoding, and an archival research status on GitHub without ongoing maintenance or immediate commercial deployment plans from Anthropic.

Key takeaways

• FlashProve uses zkSNARKs to generate cryptographic proofs confirming inference outputs from committed models without revealing weights.
• Tested on Llama 4 Maverick with a single H100, proof generation takes 70 to 95 times the GPU time of vLLM, while verification runs on standard CPUs in seconds.
• Current prototype limitations include a maximum context of 1,024 tokens, integer-quantized models, and greedy decoding, released under Apache 2.0 with Lean specifications.
• The GitHub repository is marked as an archived research artifact without external contributions, and Anthropic has announced no deployment schedule.