Anthropic's Frontier Red Team has released FlashProve, an open-source zkSNARK prototype enabling provable inference for Llama 4 Maverick, a 400B mixture-of-experts model with ~17B active parameters per token. The system allows a prover to demonstrate that a response originated from a specific committed weight set without revealing the weights, addressing risks such as model swapping or token-stream tampering in API services. FlashProve operates in two phases: an offline commitment phase taking ~39 minutes covering ~16 million requests, and an online proving phase generating cryptographic proofs via an interactive oracle proof compiled with hash functions, offering post-quantum security. On a single H100 80GB GPU with 24 CPU cores and ~290GB host RAM, FlashProve achieves ~110 tokens/second (batch of 24), producing 18.78MB proofs verified on a CPU in 17 seconds (24 threads). Proving incurs a 70-95x GPU-time overhead versus vLLM FP8 inference on 8 H100s. Context is capped at 1,024 tokens with greedy decoding. Compared to prior zkML systems (Attestable, VerInf, DeepProve, zkLLM), FlashProve offers higher throughput on larger models, though with larger proofs. Security properties—completeness, soundness, and zero-knowledge—are formalized in Lean, bounding cheating probability at ~q/2^124 + 2^-101. Two verifiers ship: a fast C++ version for production and a slower Lean-verified version with formal guarantees. The code is Apache 2.0 licensed.