ARC Prize verifies DeepSeek V4 Flash: 61% on ARC-AGI-2 at four cents a task
The ARC Prize team published verified semi-private results for DeepSeek's open-weight V4 Flash 0731: 89.0% on ARC-AGI-1 at about two cents per task, and 61.4% on ARC-AGI-2 at about four cents, at maximum reasoning effort. The model is a 284-billion-parameter mixture-of-experts that activates only 13 billion parameters per token and handles a 1-million-token context. The result shot to the top of Hacker News, where the price-performance ratio drew most of the attention.
Why it mattersARC-AGI-2 was built to resist memorization and reward genuine abstract reasoning, and a score like this from a small-activation open model at pocket-change prices closes one more gap between open and closed AI. For developers it means near-frontier reasoning is now something you can download and run — not just rent.
▶ Related video: DeepSeek V4 Flash GA IS INCREDIBLE! Powerful, Cheap, & Fast! (Fully Tested)
Read in the app — free, in 9 languages
Related stories
A model learned to write without the method that trains every AI.2026-10-06Mathematicians cracked five open problems using an ordinary chat box.
2026-10-05Untuned models solved agent tasks their polished versions could not.
2026-10-04Pretraining on everyday photos got a model to 70% on a reasoning test.
2026-10-04An AI trained for $8,000 beat Stratego's greatest player.
2026-10-03