Artificial intelligence / news brief
Anthropic open-sources FlashProve to verify which model produced an AI response
- Event
- Published

Anthropic released FlashProve, an Apache-2.0 research prototype that binds model outputs to committed weights using zero-knowledge proofs. On a quantized Llama 4 Maverick, Anthropic reports 110 proved tokens per second and CPU verification in 17 seconds, but proving used roughly 72–94 times the per-GPU time of ordinary inference.
Why it matters
Cryptographic model identity could strengthen audits and routed inference, but this prototype has short contexts, high overhead and no independent security audit.
Sources
- Progress on provable inference for large language models (external site)Anthropic · primary source ·
- Anthropic Releases FlashProve for Cryptographic AI Inference Proofs (external site)Unite.AI · coverage ·