Anthropic releases FlashProve in open source, a tool for provable inference
Anthropic published a prototype of an open cryptographic protocol called FlashProve that lets anyone verify a model response without access to model weights.
Anthropic released FlashProve in open source, a tool for provable inference.
Essentially, it is a prototype of an open cryptographic protocol through which any verifier, without access to model weights, can confirm that the response was generated by the exact model claimed by the provider.
This capability is useful, for example, to check that a routing service is not swapping a model for a cheaper one, or that a model inside a company’s environment has not been replaced by attackers.
The mechanism works as follows: the provider publishes a digital fingerprint of its weights in advance, and then attaches a proof file to each response. A match between that file and the fingerprint confirms the model’s authenticity, and the check can be run on an ordinary computer without a GPU or the weights. The proof is about 19 MB per 1,024-token request, and verification takes about 17 seconds. This is still much more expensive than normal inference, but for comparison, the previous result required about 14 hours of verification and a 93 GB proof.
The algorithm is based on zkSNARKs. The method has been known for a long time, but Anthropic’s contribution is that it optimized it to practical speeds for models with up to 400B parameters. In the Anthropic research article, it says that Claude mainly worked on the acceleration and optimization.
This may be a step toward detectors of text provenance.
Why it matters
AnalysisProvable inference can make model routing and enterprise deployments auditable, reducing the risk of silent model substitution.
Discuss in community
Share your questions and insights with developers
Related in Ecosystem
Practical guides, core terminology, and agent skills
Anthropic introduces FlashProve prototype for provable AI inference
Anthropic has presented FlashProve, a prototype system that lets developers mathematically verify that an answer came from a specific AI model without exposing the model weights.
AI Radar: GPT-6.1 Terra, Gemini 4 Carbon, Mistral Large 4 and OpenAI/Anthropic pricing changes
A roundup of recent AI news: new LLM releases, OpenAI and Anthropic subscription changes, world models, video generation, and research updates.
Claude Haiku (Fast Compact Models by Anthropic)
The most compact and fastest model in the Anthropic lineup, designed for lightning-fast responses, mass text classification, data extraction, and agent routing at minimal cost.
Claude Opus (Flagship AI Model by Anthropic)
The most powerful heavyweight model in the Anthropic family, designed for complex philosophical and scientific analysis, multi-layered text generation, and deep contextual understanding.