Anthropic's Flashprove prototype pays a much larger computational premium for a different property: proving that an inference calculation was performed with a committed model. Its integer-quantized approach uses a cryptographic proof, a zk-SNARK. The reported single-H-one-hundred rate is about a hundred and ten tokens a second, with a nineteen-megabyte proof per thousand and twenty-four tokens. GPU time is seventy to ninety-five times the vLLM baseline. That proof concerns the computation used to produce the tokens; it doesn't certify that the answer is true. It is an older research prototype, with a substantial proving cost still attached.

Source 1