• Sources: primary, discussion
  • Summary: ARC Prize published verified semi-private scores for DeepSeek V4 Flash 0731. At max effort it scores 89.0 percent on ARC-AGI-1 at 0.02 dollars per task and 61.4 percent on ARC-AGI-2 at 0.04 dollars per task. The high and low reasoning variants score 87.0 and 84.0 percent on ARC-AGI-1 and 56.0 and 46.0 percent on ARC-AGI-2. ARC Prize publishes per-task cost only for the max-effort run.
  • Why it matters: A third-party verified score paired with a per-task cost for the max-effort run gives a measured price point for reasoning-heavy agent work instead of a vendor claim.
  • Follow-up: Record any DeepSeek or investor statement on the suspended funding round tracked in f-0004.

send feedback on this story