10,000-agent swarm solved Navier-Stokes ~8 months ahead of a single GPT-6-Astra, at 48,000x token cost
alvelda · x · 2026-09-24
- scaling01 estimates the 10,000-agent swarm that solved the Navier-Stokes problem was about 8 months (range 5.5-12.4) of frontier progress ahead of a single GPT-6-Astra instance.
- The cost: roughly 48,000x more output tokens than the largest single-model evaluation, and 68x more than an entire long-horizon coding benchmark.
- alvelda adds that AI progress is so fast even experts struggle to grasp it, with profound implications, opportunities and risks.
More from Models
- Researchers call Anthropic's hyped biology paper underwhelming without the PR spin — soumitrashukla9 · 2026-09-24
- Hands-On: Opus 5.5 Impresses With 3D Scenes, Computer Use and Video Editing in Early Reviews — EricBuess · 2026-09-24
- Early user praise for Opus 5.5: stable behavior, cheaper pricing, 'Anthropic is back' — almmaasoglu · 2026-09-24
- New Paper: AI Agents Infer Your Wealth From Emails and Recommend Pricier Options — niloofar_mire · 2026-09-24
- Using Sol and Fable as sounding boards: Sol concedes to pushback far more often — dreamwieber · 2026-09-24
- Anthropic tip: make Claude actually look at DNA sequences, not just run bioinformatics tools — Sauers_ · 2026-09-24