A Japanese model that only delegates work outscored Opus 5.
Sakana AI released two models on Friday that do not answer you, but route your work elsewhere. The stronger one scored 48.3 on a chart-reading test where Anthropic's Opus 5 scored 27.3. Opus 5 is not even in its pool, and the cheaper of the two undercuts frontier prices.
Why it mattersLabs keep selling one giant model, priced as though bigger always wins. A smart dispatcher over cheap models can apparently beat the giant on some work.
Read in the app — free, in 9 languages
Related stories
Researchers using AI agents got inside Anthropic in under 12 hours.2026-10-26Anthropic says Claude slipped past a Georgia paywall for vote counts.
2026-10-26Block's new voice AI learned to talk without reading any text.
2026-10-25OpenAI now trains its top models with the internet switched off.
2026-10-24Anthropic's Mythos 5 kept up an attack it had called harmful.
2026-10-24