Anthropic published FlashProve as an open-source system for what researchers call provable inference, a cryptographic proof that a given output came from a specific model. The blog post went up Tuesday, November 3, 2026.

On Llama 4 Maverick at 4-bit quantization, the system produces roughly 19 MB of proof per 1,024-token request. Anthropic's own post calls the result intermediate progress.

The fingerprint mechanism hides weights but not architecture, so the model family remains identifiable from the proof structure. A provider matching vLLM throughput with proofs attached needs seventy to ninety-five H100s per unit of inference capacity.

Lean 4 proves the soundness bound of q/2^124 + 2^-101, which is a theoretical guarantee rather than a field test.

No document in the packet measures FlashProve on any model other than Llama 4 Maverick, and none states whether the system has been tested against Claude weights. The technical report PDF and the full Lean statements were not readable in the packet.

The GitHub repository is tagged not maintained and not accepting contributions at version 0.16.0, and no lab plans to deploy it.

The draft buries the constraint under the headline number and repeats the 70-to-95x figure twice in adjacent paragraphs. The rewrite leads with the overhead, consolidates the throughput spec into a single supporting line, and ends on the maintenance status so the reader leaves with the deployment verdict rather than a restated multiplier.