The number that stopped me wasn’t 89.0% on ARC-AGI-1. It was $0.02 per task. I’ve spent the last year writing prompts as if every call were expensive — batching, caching, second-guessing whether a job was worth the spend. At two cents for ARC-AGI-1 and four for ARC-AGI-2, that whole reflex is obsolete for anything I’d run at my scale.
The more useful detail is the three reasoning variants. Low gets 84.0% on ARC-AGI-1 and 46.0% on ARC-AGI-2; Max gets 89.0% and 61.4%. That gap is a dial, and the ARC-AGI-2 end of it is where the dial actually earns its keep — fifteen points between Low and Max, versus five on ARC-AGI-1. So the design question for anything I build is which tier a task deserves, not which model. Cheap tier for the boring pass, expensive tier for the one that has to be right.
What I’d do: stop treating reasoning effort as a global setting and put it in the job definition. Route per task, log which tier solved what, and let the log tell you where you overpaid.
The story — DeepSeek published ARC-AGI results for V4 Flash 0731 on Jul 31, 2026, with three reasoning variants. At max effort it scores 89.0% on ARC-AGI-1 Semi-Private at $0.02 per task and 61.4% on ARC-AGI-2 Semi-Private at $0.04 per task; the High variant scores 87.0% and 56.0%, and Low scores 84.0% and 46.0%. (Source)