Anthropic’s Frontier Red Team released FlashProve, an open-source (Apache 2.0) prototype for “provable inference.” It uses zkSNARK proofs so you can check that a reply really came from a specific, committed model.
The numbers
- Tested on Llama 4 Maverick (400B MoE, about 17B active per token)
- About 39 minutes once per model to build the commitment
- Proofs at about 110 tokens/sec on one H100, at roughly 70-95x the GPU time of normal vLLM inference
- Proofs around 19 MB for a 1,024-token request, checkable on a CPU without the weights
The catches
- It proves an integer version of the weights with greedy decoding, not normal sampled output
- Context is capped at 1,024 tokens
- The protocol is proved in Lean, but the C++ verifier and CUDA prover are only tested
- No plans to deploy it at Anthropic or anywhere else yet
My take
This solves a real problem. When you pay for a model through an API or a reseller, you’re trusting they didn’t quietly swap in something cheaper. Way too slow for everyday use today, but it’s a real step.
God Bless, and Use the Source!
