{"slug": "anthropic-s-flashprove-proves-which-model-wrote-a-reply-at-110-tokens-s", "title": "Anthropic's FlashProve proves which model wrote a reply at 110 tokens/s", "summary": "Anthropic's FlashProve system lets a customer verify which model produced a reply using a zero-knowledge proof, at a throughput of 110 tokens per second, without exposing model weights. The verification carries roughly 70-95x the GPU cost of plain inference, according to the report, addressing a gap where customers previously had to trust a router or provider that the model they paid for answered.", "body_md": "Until now a customer had to trust a router or provider that the model they paid for answered. A zero-knowledge proof makes that checkable without exposing weights, at roughly 70-95x the GPU cost of plain inference.\nRead: Until now a customer had to trust a router or provider that the model they paid for answered. A zero-knowledge proof makes that checkable without exposing weights, at roughly 70-95x the GPU cost of plain inference.\nRead: A mature database is now finding old bugs faster than its committers can fix them, and the cost is a late major release with large features reverted. Projects adopting AI bug finding inherit that patching load.\nRead: Astra fixed Qwen3.5-0.8B reasoning loops by editing 0.016% of its weights, with no training run. Model repair has meant gradient updates and eval cycles; this is a direct edit, and the full experiment logs are public for others to check.\nRead: Clients send only new messages plus the previous message ID instead of the full history each turn. Long-running agents stop re-uploading growing transcripts, and Claude begins responding sooner because the server already holds the state.\nRead: OpenAI is pairing the higher bounties with an internal sprint where Cathedral Systems researchers use its models against its own defenses, and with wider model access for trusted defenders. Frontier models are becoming a standard tool on the defensive side.\nRead: Run time also fell from 11 minutes to about 2, and quality rose 44% on Datadog's internal benchmark. At $25 per alert, running an agent on every page was expensive; at about $4 the per-alert cost drops by more than 80%.\nRead: An agent building a payment integration gets working test API keys with no registration, and a person claims the account later. Signup was the step that stopped an agent and waited for a human; Stripe built this with Vercel, Replit and Lovable.", "url": "https://wpnews.pro/news/anthropic-s-flashprove-proves-which-model-wrote-a-reply-at-110-tokens-s", "canonical_source": "https://www.vibeleaderboard.ai/intel/brief/2026-11-03", "published_at": "2026-11-03 12:10:49+00:00", "updated_at": "2026-11-03 12:53:49.417334+00:00", "lang": "en", "topics": ["ai-safety", "large-language-models", "artificial-intelligence", "ai-infrastructure"], "entities": ["Anthropic", "FlashProve"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/anthropic-s-flashprove-proves-which-model-wrote-a-reply-at-110-tokens-s", "markdown": "https://wpnews.pro/news/anthropic-s-flashprove-proves-which-model-wrote-a-reply-at-110-tokens-s.md", "text": "https://wpnews.pro/news/anthropic-s-flashprove-proves-which-model-wrote-a-reply-at-110-tokens-s.txt", "jsonld": "https://wpnews.pro/news/anthropic-s-flashprove-proves-which-model-wrote-a-reply-at-110-tokens-s.jsonld"}}