Anthropic FlashProve builds on StarkWare-co-invented STARKs

November 3, 2026

Anthropic has released FlashProve, an open-source research prototype that cryptographically proves which AI model generated a reply, and StarkWare notes the system is built on STARK cryptography its founders co-invented. StarkWare touted the move via Starknet's official X account as a signal for verifiable AI.

FlashProve attaches a zero-knowledge proof to each response confirming it came from a specific committed model, without revealing the model's weights.
In tests on Meta's 400B-parameter Llama 4 Maverick, proof generation ran at roughly 110 tokens/second on a single H100 — about 70–95× the GPU time of plain vLLM inference, with proofs around 19 MB each.
The prototype is open source, aimed at letting users verify model provenance in AI responses.

FlashProve attaches a zero-knowledge proof to each response confirming it came from a specific committed model, without revealing the model's weights.
In tests on Meta's 400B-parameter Llama 4 Maverick, proof generation ran at roughly 110 tokens/second on a single H100 — about 70–95× the GPU time of plain vLLM inference, with proofs around 19 MB each.
The prototype is open source, aimed at letting users verify model provenance in AI responses.

Anthropic FlashProve builds on StarkWare-co-invented STARKs

Sources

Anthropic FlashProve builds on StarkWare-co-invented STARKs

Sources