Paying top-tier prices but getting a watered-down model was once the most hidden gray area in the large-model supply chain. Anthropic has teamed up with zero-knowledge proof researchers to open-source FlashProve.

Verification has three steps: before launch, it takes about 39 minutes on a single H100 to generate a "fingerprint" ID for the weights and make it public; after each conversation, the GPU generates a proof of about 19 MB; users run cp_verify on an ordinary CPU and determine authenticity in 17 seconds, without needing the weights or a GPU. The underlying zkSNARK performs about 2,000 sumchecks per request, and security relies only on the BLAKE2 hash.

AI writes the code, Lean serves as judge, and humans uphold only six ironclad rules

The hard part is compute: the test used Llama 4 Maverick with 400 billion parameters, whose 4-bit weights reach 232 GB, while a single H100 has only 80 GB of VRAM. FlashProve leverages MoE—only about 17 billion parameters are activated per token, so the proof size is tied only to those 17 billion and skips 95% of the non-active parameters, while the routing is hidden in an indicator vector for secrecy; the weights are kept in memory and streamed into the GPU in segments to swallow the giant. Even more hardcore is the repository: 95.2% of the code (about 398 MB) is Lean 4 mathematical proofs, and the 114-page paper and the underlying code were almost entirely produced by Claude.

Humans only finalize six security definitions with AI—completeness (false rejection below one in a trillion), knowledge soundness (forgery probability below one in 2 to the 100th power), and zero-knowledge (repeated verification does not leak the weights). Unable to read through the massive derivations, humans set the axioms, AI reasons, and mathematics gives the final verdict, forming a safety contract on the path to superintelligence.