Anthropic has announced the development of an open-source system called FlashProve, which provides cryptographic confirmation of the origin of artificial intelligence responses. According to the developers, the problem is systemic: an ordinary user cannot verify which specific language model generated the received response. This opens up the possibility for routing services or the developers themselves to secretly substitute expensive models with cheaper ones to cut costs. The new system aims to close this vulnerability by turning the "origin" of a response from an unprovable claim into a verifiable cryptographic fact.
Why Text Quality Assessment is Unreliable
Anthropic emphasizes that assessing results solely by text quality remains an unreliable method of control. Even surrogate or compromised systems are capable of producing outwardly plausible results while hiding the actual generation process from the end consumer. In an environment where models of varying classes and costs operate simultaneously, and corporate clients increasingly combine multiple systems, the visual "plausibility" of a response ceases to be proof of which specific model created it. That is why the company is betting on cryptographic inference proofs rather than heuristic quality evaluation.
Two Phases: Weight Fingerprint and On-Demand Proof
The FlashProve architecture relies on cryptographic inference proofs and consists of two main phases for reliable control. In the offline phase, the system publishes a short digital fingerprint of the model's weights without revealing proprietary parameters. In the online phase, a separate proof is created for each specific request and generated tokens. The resulting file can be verified by the user or an independent auditor on a regular computer in a matter of seconds, which developers claim completely eliminates the possibility of covertly falsifying results.
Performance and Open Source
When tested on the Llama 4 Maverick model with 400 billion parameters, the system demonstrated high performance: a single H100 graphics processor generates confirmations at a rate of about 110 tokens per second. A proof file of 1024 tokens takes up 19 MB, and its verification on a regular central processing unit takes just a few seconds, allowing the technology to be used without excessive computational overhead. Anthropic has published the technical specification and working code of the system on GitHub under an open license, counting on active community participation in its improvement.
Context: Price War in the AI Industry
The announcement comes at a time when competition in artificial intelligence has turned into a price war. According to 3dnews.ru, Anthropic broke ahead this year thanks to advanced models and a coding service highly praised by American business, but OpenAI has narrowed the gap in the corporate segment by introducing the GPT-5.6 lineup in Sol, Terra, and Luna variants with various capabilities and more economical options. Rising business spending on AI and the realization that not all tasks require the most powerful models have led companies to increasingly turn to low-cost systems with open weights. Against this backdrop, the possibility of covertly swapping an expensive model for a cheap one becomes not a theoretical, but a practical threat, which FlashProve addresses.
Significance for Trust and Regulation
Anthropic is convinced that introducing such solutions significantly increases trust in artificial intelligence. The technology, according to the company, also opens up new opportunities for industry regulation, safety verification, and compliance monitoring by developers. For independent auditors and regulators, the ability to verify on a regular computer in seconds which model generated a response turns oversight from a subjective assessment into a formalizable procedure.
