Anthropic releases FlashProve, a zkSNARK prototype for verifying AI model outputs
In brief
- Anthropic unveiled FlashProve, an open-source research prototype, on November 2, 2026.
- FlashProve uses a hash-based zkSNARK to verify LLM outputs without revealing model weights.
- Throughput reached about 110 tokens per second on one NVIDIA H100 running Llama 4 Maverick, per Anthropic.
- Proving overhead ran roughly 70 to 95 times standard vLLM inference GPU time.
- Anthropic says FlashProve is not ready for production deployment.
What FlashProve is trying to fix
Right now, users can't confirm whether a response came from the model a developer claims or from a cheaper one swapped in by a routing service, Crypto Briefing reported. FlashProve is Anthropic's cryptographic answer to that.
A zkSNARK lets one party show a computation was performed correctly without revealing the private inputs behind it. Here's how FlashProve uses one. The model provider first commits to a set of weights in an offline step (Anthropic says the commitment is about 6.5 MB and takes roughly 39 minutes to generate). Every inference after that produces a succinct proof confirming the committed model generated the claimed output tokens for a given prompt, while the weights, activations and routing decisions stay confidential.
That's the big departure from trusted-hardware approaches, since FlashProve relies on math rather than a secure chip to vouch for the computation.
Anthropic's test numbers (and the catch)
Anthropic tested the system on Meta's 4-bit quantized Llama 4 Maverick, a 400.7-billion-parameter model. The figures below are the company's own reported results. Nobody has independently verified them.
On a single NVIDIA H100 GPU, FlashProve hit around 110 tokens per second and processed a batch of 24 requests in roughly 223 seconds. Each 1,024-token request produced a proof of about 19 MB. A standard CPU can verify one in 17 to 21 seconds, with no access to the weights.
The cost is steep. Proving overhead runs at roughly 70 to 95 times the GPU time of standard vLLM inference, and Anthropic is upfront about where that leaves the project.
Anthropic labels FlashProve as an intermediary research tool, not something ready for production deployment.
Who built it
Roy Rinberg and Alex Allain of Anthropic led the project, according to the published materials as reported by Crypto Briefing. Contributors included Alessandro Chiesa, Giacomo Fenzi and Guy Weissenberg from EPFL (Chiesa is also affiliated with StarkWare), plus Miranda Christ of MIT and RESI, Rahul Dalal of Anthropic's STEM Fellows Program and Succinct, and Ben Mayer of the Anthropic Fellows Program and MIT.
The proofs' formal security properties were machine-checked in Lean, a proof assistant. All code and technical materials are available under the Apache 2.0 license on the anthropics/FlashProve GitHub repository.
Frequently asked questions
What problem does Anthropic's FlashProve try to solve?
Users can't currently confirm whether a response came from the model a developer claims or from a cheaper one substituted by a routing service. FlashProve generates a cryptographic proof that a committed model produced the claimed output tokens for a given prompt.
How does FlashProve verify outputs without revealing model weights?
The provider first commits to a set of weights in an offline step, a commitment of about 6.5 MB that takes roughly 39 minutes to generate. Each inference then produces a succinct zkSNARK proof, while the weights, activations and routing decisions stay confidential.
Is FlashProve ready for production use?
No. Anthropic labels FlashProve an intermediary research tool that isn't ready for production deployment. In Anthropic's reported tests, proving overhead ran at roughly 70 to 95 times the GPU time of standard vLLM inference.
