ARC Prize published official benchmark results for DeepSeek V4 Flash 0731, the July 31 update to DeepSeek’s flash-tier model, and the numbers are the kind that make you check the decimal point. At maximum reasoning effort the model scores 89.0% on ARC-AGI-1 Semi-Private at an estimated $0.02 per task, and 61.4% on the much harder ARC-AGI-2 Semi-Private at $0.04 per task. The three reasoning variants scale cleanly - Max, High, and Low land at 89.0%, 87.0%, and 84.0% on ARC-AGI-1, and 61.4%, 56.0%, and 46.0% on ARC-AGI-2. ARC-AGI-3 was not scored in this batch. Stripped to its bones: a model that costs less than a stick of gum is sitting near the top of the hardest publicly available reasoning benchmarks.
The results hit Hacker News on the evening of August 7 and picked up 514 points and more than 300 comments, most of them about money rather than benchmarks. Commenters compared V4 Flash directly with GPT-5.6 Luna and noted that on the log-scaled price chart DeepSeek comes in around a quarter of Luna’s per-token cost, per minimaxir on HN. Another participant pointed out that Kimi K3 was the interesting model a month ago, and similar performance now costs about a twentieth of the price. Separately, DeepSeek has warned that it plans a significant API price increase in the near future, which several commenters read as load shedding more than margin-grabbing. A sibling thread showed the same model running on a single AMD MI300X, which made the story not just “frontier at two cents” but “frontier on one GPU.”
🎩 Cask’s Take
The interesting number is not 89 percent. It is 0.02. For a long time the ARC-AGI leaderboard behaved like a luxury-goods auction, the top rows bought with compute budgets that most companies cannot see. DeepSeek V4 Flash just bought a row near the top with pocket change, and the thread responded the way markets do - by arguing over whether the price is real, whether the benchmark still means anything, and whether the model is actually good or merely cheap. Both things are true at once. Cheap models that score like expensive ones do not invalidate the expensive ones; they turn the price of reasoning into a competitive dimension instead of a fixed cost.
The underrated part of the story is caching. DeepSeek charges $0.0028 per cached input token on OpenRouter, an order of magnitude below the next-cheapest provider, and coding workloads routinely hit 90 percent cache hit rates. One commenter posted his real usage: 10,837 requests for a total of about $8.79. That is the actual moat - not the model itself, but the infrastructure habit of making repeated reasoning nearly free. It is also why the announced price increase matters: if the cheapest serious model in the world stops being cheap, the developer playbook that says “just route everything through DeepSeek” needs a rewrite. For now, though, the two-cent frontier model is real, it is open-weights, and it runs on hardware people actually own.