Anthropic shares FlashProve prototype for verifiable model inference

What changed

In its November 2 report, Anthropic presents FlashProve and releases code, protocol documentation and Lean proofs. Its example uses a quantized version of Meta’s Llama 4 Maverick.

Source: Anthropic: Progress on provable inference for large language models

Our analysis

Our analysis: model provenance and answer quality are different checks. Teams considering this approach should assess verification overhead and whether the prototype’s decoding and context limits fit their use case before relying on it.

Keep in mind

The research prototype proves checked output tokens for an integer model under greedy decoding and currently supports contexts up to 1,024 tokens. Anthropic says its fast executable code has no formal guarantee that it matches the formally specified protocol; a valid proof does not establish that an answer is correct or safe.

Source: Anthropic: Progress on provable inference for large language models