Anthropic open-sources FlashProve: provable inference costs ~165x more than raw inference
arthurcolle · x · 2026-11-03
Anthropic released FlashProve, an open-source library for provable inference, benchmarked on Llama 4 Maverick (a 400B MoE). @0xOsprey ran the economics:
- Huge gap: a 1,024-token request costs $0.0001 on stock vLLM vs $0.017 proven — roughly 165x; throughput drops from 37.2M to 0.4M tok/hr per GPU, latency goes from seconds to 4-5 minutes
- Heavier prover hardware: 1×H100 + 24 CPUs + 280GB RAM, plus an 18.78MB proof payload
- Verification is dirt cheap: client-side CPU-only checks (21s/$0.0003 via C++, 24min via Lean machine-checked soundness)
- Bottom line: proving is the bill; economics only work as client-sampled auditing on 1% of traffic (+17% cost), full proving is impractical
Related event: Anthropic Open-Sources FlashProve for Verifiable AI Reasoning(2 posts)→
More from Infra
- 500M-row benchmark: hand-written BFS beats NebulaGraph in MySQL, 2.7ms vs 23.6ms — JeremyCMorgan · 2026-11-03
- PostHog re-rebuilds its data warehouse five months after switching from ClickHouse to DuckDB — teodorio · 2026-11-03
- Local MoE inference on a Tesla V100: 55 tok/s decode with 8,109 experts cached in 31.6GB VRAM — Traditional_Bell8153 · 2026-11-03
- Cities Worldwide Build AI Services on NVIDIA Nemotron, Cosmos and Omniverse — NVIDIAAI · 2026-11-03
- Anthropic spends $2.4B ARR on data vendors — just 6% of its compute bill — himanshustwts · 2026-11-03
- StoneX sees $309B in free cash flow losses for hyperscalers and neoclouds next year — GaryMarcus · 2026-11-03