Anthropic's Frontier Red Team has released FlashProve, an open-source research prototype that applies zkSNARK zero-knowledge proofs — the cryptography behind blockchain scaling — to large-model inference verification, according to Anthropic's technical blog and multiple outlets on November 3. It lets users verify that a response truly came from a specified model without accessing its weights or re-running the model.
The problem it targets sits on the buyer's side: when a model is resold, distilled or deployed opaquely, customers struggle to confirm the API endpoint actually runs the model they pay for. FlashProve constructs hash-based zkSNARK proofs that a given output was produced by specified weights, letting vendors prove authenticity without disclosing weights.
The work was co-authored with external researchers including StarkWare co-founder Alessandro Chiesa, and Anthropic says Claude helped improve the system. Reported benchmarks: on the 400-billion-parameter Llama 4 Maverick, a single H100 sustains roughly 110 tokens per second of proven inference, at a cost of 70 to 95 times ordinary inference.
The technique addresses trust in the model supply chain. When inference services are resold, distilled or deployed in opaque environments, the mapping between weights and service is hard to audit; cryptographic proof offers a path that does not rely on vendor self-discipline. Blockchains use zkSNARKs to prove transactions valid without revealing their details — FlashProve applies the same lineage of techniques.
Production use remains distant: the 70-95x overhead suits spot audits rather than per-query verification, and full proof latency and cluster requirements are still undisclosed. But in model marketplaces, compliance auditing and tiered inference markets, verifiable inference has moved from paper concept to open prototype. The signal matters: trust in the model market is starting to shift from reputation to cryptography.