DeepSeek V4 Flash 0731 turns ARC-AGI into a cost-per-task argument

DeepSeek V4 Flash 0731 landed on the ARC Prize results page with numbers that are hard to ignore. ARC Prize lists the Max effort variant at 89.0% on ARC-AGI-1 Semi-Private and 61.4% on ARC-AGI-2 Semi-Private. The same page shows estimated costs of $0.02 per task for ARC-AGI-1 and $0.04 per task for ARC-AGI-2. Its High and Low reasoning variants are also reported: High reaches 87.0% and 56.0%, while Low reaches 84.0% and 46.0%.

The interesting part is not only the score. ARC-AGI-2 is treated as a harder test of adaptation to novel tasks, so a strong result at a low task cost changes the practical question. Instead of asking only which model is most capable, teams can ask how many reasoning passes they can afford to run inside a coding agent, data-analysis loop, or document workflow.

The Hacker News thread reflected that shift. The submission, posted at 2026-08-07 17:56:20Z, had 718 points and 432 comments at crawl time. Discussion quickly moved from benchmark position to everyday use: coding speed, cache behavior, long-running sessions, and whether the model stays reliable when tools are involved. Some practitioners described it as useful for debugging and document analysis; others reported loops, odd drift, or failed tool behavior.

That mix matters. A low-cost reasoning model can make agentic workflows feel less constrained, but only if it remains steady over long tasks. Cheap tokens do not help much when a model repeats a bad plan, talks instead of calling a tool, or requires human cleanup after every few steps. The benchmark result opens the door; operational reliability decides how wide it stays open.

What the ARC Prize table makes clear is that model competition is moving into cost-per-task territory. Frontier-style scores still matter, but agent builders also care about throughput, cache hit rates, and the price of trying again. HN’s reaction was valuable evidence of that migration: the community was not only ranking DeepSeek V4 Flash 0731, it was testing whether the model belongs in the daily-driver stack.