AI
Anthropic's FlashProve Lets You Verify AI Model Identity—but at a Steep Speed Cost
Cryptographic proofs can now confirm which AI model generated your response, but the computational overhead makes it impractical for everyday API calls.
1/2
The New Stack
When you call an AI API, you have no reliable way to confirm the model serving your request is the one you paid for—a provider could quietly swap in a cheaper alternative or an attacker could tamper with responses. Anthropic is taking aim at this trust problem with FlashProve, a cryptographic verification system that proves which AI model produced your output without revealing its internal weights.
Released this week as open-source research code, the system represents a meaningful advance in AI accountability, but it comes with a catch that limits its immediate usefulness.
How the proof works#
FlashProve employs zero-knowledge cryptography (zkSNARKs) to mathematically establish that a response originated from a specific set of model weights—while keeping those weights private. The verification chain only holds if the initial model commitment comes from a trusted party.
Anthropic tested the system on Meta's Llama 4 Maverick, a 400-billion-parameter mixture-of-experts model. The company formalized the underlying protocol using Lean, a proof assistant, to verify that valid proofs always verify and forged proofs fail with overwhelming probability. The actual verification runs fast: proofs shrink to roughly 19 MB and verify in seconds on a single CPU thread.
The performance penalty#
The real problem emerges on the generation side. To create proofs, FlashProve demands 70 to 95 times more GPU time than standard inference with vLLM. Processing occurs in batches after requests complete, not in real-time as responses stream in—unsuitable for applications requiring immediate cryptographic guarantees.
On an Nvidia H100 GPU, the system processed approximately 110 prompt tokens per second while building proofs. For context, an earlier system called VerInf required 14.3 hours to generate a single proof covering 1,000 tokens, with a 93.6 GB proof file taking nearly a day to verify. FlashProve is dramatically faster, yet still expensive compared to ordinary inference.
For teams already scrutinizing GPU costs token-by-token, absorbing nearly 100× overhead would be difficult to justify in production workflows.
Current limitations and the research-only status#
FlashProve operates under constraints that keep it squarely in the research phase. It supports only 1,024 tokens of context, relies on integer-based inference instead of floating-point arithmetic, and uses greedy decoding rather than the sampling methods production models employ.

