Clarity for AI buildersSponsored by AI Routing
availability

Anthropic’s FlashProve prototype verifies model outputs without revealing weights

Anthropic’s FlashProve prototype verifies model outputs without revealing weights
Anthropic has released FlashProve, an open-source research prototype designed to cryptographically verify that an AI response came from a specified model without revealing its weights, according to The Decoder. The system addresses the risk of a provider or routing service substituting another model, but its context and sampling limits keep it from covering typical chatbot use. FlashProve uses a cryptographic technique called zkSNARK. A provider first creates a digital fingerprint of a model’s weights, then supplies a proof with each response showing that the specified model generated the output. The proof conceals the weights, though it does not hide the model’s architecture. According to the technical report described by The Decoder, FlashProve generates proofs at roughly 110 tokens per second on a single H100 GPU for Meta’s Llama 4 Maverick, a 400-billion-parameter model. Each proof is about 19 megabytes and takes 17 seconds to verify on a standard CPU. Proof generation adds roughly 70 to 95 times the computational overhead of normal inference. That remains substantial overhead, but Anthropic reports a large improvement over the only comparable open-source system previously available. According to the company, that system took more than 14 hours to process 1,000 tokens, produced a 93.6-gigabyte proof and required nearly a full day for verification. Anthropic proposes several uses for inference verification. Regulators could check whether providers are running approved models, companies could demonstrate that a specified safety model reviewed their code, and providers could detect unauthorized changes to computations on their servers. These are proposed applications, rather than demonstrated deployments in the supplied report. The company’s Frontier Safety Roadmap set a goal of building an inference-verification prototype by September 2026. Anthropic says it had already met that target through a separate project that signs inference traffic and traces computation on a trusted server. FlashProve takes a different approach, relying on cryptography rather than trusted hardware. The prototype has significant restrictions. It currently proves computations only for contexts of up to 1,024 tokens. It also works only when the model selects the single most likely next token, rather than using the randomized sampling common in chatbot responses. Those limits narrow what the system can verify today, despite its improved proof-generation performance. The Decoder reports that FlashProve’s full code, Lean proofs and documentation are available on GitHub under the Apache 2.0 license.