All AI news

Anthropic releases FlashProve: model provenance is not answer correctness

The research prototype supplies inference proofs, not a new Claude web or API feature.

Anthropic released FlashProve and Apache 2.0 code on November 2. The research prototype aims to prove that checked tokens match a committed integer model’s greedy output without revealing its weights. Exact-match guarantees do not cover unchecked portions; the commitment ID must still come from a trusted source. This does not establish answer accuracy or safety.

The prototype covers a specific integer model, greedy decoding and contexts up to 1,024 tokens. Fast executable code is not guaranteed to match the formal proof; the repository says it is not maintained. Full inference proving requires a specified Linux/H100 environment; compute still incurs costs, and third-party model weights have separate license terms.

This research release is not a Claude web, mobile-app or commercial API rollout. HK, Macau and Taiwan deployment eligibility and Traditional Chinese/Cantonese interfaces are unconfirmed. No new subscription price was announced. The release time is unspecified. Checked November 4 UTC; not installed or tested.

Usage conditions

Hong Kong
Hong Kong: local deployment eligibility and Traditional Chinese/Cantonese interfaces are unconfirmed; this is not a Claude service rollout.
Macau
Macau: local deployment eligibility and Traditional Chinese/Cantonese interfaces are unconfirmed; this is not a Claude service rollout.
Taiwan
Taiwan: local deployment eligibility and Traditional Chinese/Cantonese interfaces are unconfirmed; this is not a Claude service rollout.

Official source: Progress on provable inference for large language models

Source published:

Source checked:

Published here: