ARC Prize Publishes DeepSeek V4.1 Flash Scores on ARC-AGI (Verified)
Key Highlights
ARC Prize published DeepSeek V4.1 Flash results on ARC-AGI (Verified): 72.9% on ARC-AGI-2 at 0.13 dollars per task, and 94.5% on ARC-AGI-1 at 0.07 dollars per task. Versus V4 Flash's best, V4.1 raises ARC-AGI-1 by 5.5 points and ARC-AGI-2 by 11.5 points, though per-task cost rises about 250%, a classic capability-versus-cost trade.
What It Does
ARC-AGI tests few-shot abstract reasoning, the ability to infer a transformation rule from a handful of examples and apply it to a new one, which is a proxy for genuine generalization rather than memorization. DeepSeek V4.1 Flash's gains on both ARC-AGI-1 and the harder ARC-AGI-2 indicate improved low-sample reasoning, and the sub-dollar per-task cost keeps it practical for volume use even with the 250% increase over the prior version.
Technical Details
The per-task cost rising 250% while scores climb suggests the improvement comes partly from more inference compute, either more sampling or longer reasoning, which is the familiar pattern of trading money for accuracy. ARC-AGI-2 at 72.9% remains well below the near-perfect ARC-AGI-3 scores being posted by frontier labs, so DeepSeek V4.1 Flash sits in the strong-but-not-frontier tier on the hardest variant, while leading comfortably on the easier ones.
Versus Alternatives
Against frontier models posting 90%+ on ARC-AGI-3, DeepSeek V4.1 Flash is clearly behind on the hardest benchmark, but it offers a different value: an open-weight model with strong ARC-AGI-1 and competitive ARC-AGI-2 performance at very low cost. Against closed APIs, its open weights mean self-hosting and data control, and its Flash efficiency family has long targeted cheap, high-volume deployment.
Industry Impact and Use Cases
Strong few-shot reasoning at low cost is exactly what enterprises want for tasks that require adapting to new rules from few examples, such as novel format parsing, rare-case diagnosis, and rapid domain adaptation. DeepSeek's continued gains on ARC-AGI reinforce the trend of open-weight models closing the gap on reasoning, giving teams a credible, affordable alternative to frontier closed models for many real workloads.
For engineers, the 250% cost increase versus V4 Flash is the number to weigh against the accuracy gain, because in production the question is almost never top score but best score-per-dollar for your volume. DeepSeek's open weights mean you can run this yourself and keep data in-house, which for many regulated workloads is the deciding factor regardless of a few points on a benchmark. The practical plan is to A/B it against your current model on real few-shot tasks, watching both quality and the true per-task cost at scale. If it wins on your workloads, it is a rare combination of open, cheap, and genuinely strong on reasoning.
The bigger picture is that open-weight models are steadily eroding the reasoning advantage once reserved for closed frontrunners, and DeepSeek's ARC-AGI gains are one more data point in that trend. For organizations, the strategic implication is that betting everything on a single closed provider is increasingly unnecessary, because strong, self-hostable alternatives now cover many reasoning-heavy workloads at a fraction of the cost. The cost increase versus the prior version is a reminder that efficiency still matters, but the absolute per-task price remains low enough for volume use. The net effect is more optionality and less lock-in, which is good for everyone except vendors whose pricing depended on a capability moat.
The closing thought is that optionality is the real prize, and open-weight models like DeepSeek V4.1 Flash expand it by bringing strong reasoning within reach of teams that cannot or will not depend on a closed provider. The cost increase versus the prior version is modest in absolute terms and worth it for the accuracy gain on hard few-shot tasks, but it should still be measured against your volume. The healthy dynamic is competition on both capability and price, and every open-weight gain narrows the gap that closed vendors once used to justify premium pricing, which benefits builders far more than any single benchmark.
The DeepSeek result is notable for pairing strong reasoning with remarkably low cost. Scoring 72.9 percent on ARC-AGI-2 at about thirteen cents per task, and 94.5 percent on ARC-AGI-1 at roughly seven cents, puts capable fluid reasoning within reach of budget-constrained teams and hobbyists alike. The cost-per-task framing is the right one, because it reframes these models from research curiosities into practical components you can call thousands of times a day. DeepSeek has repeatedly shown that aggressive engineering and open weights can match or approach closed-frontier quality at a fraction of the price, and each such release tightens the competitive pressure on vendors who charge premiums. The honest caveat is that ARC-AGI scores measure a specific kind of few-shot reasoning, not the full range of production needs like long-context stability or tool use, so teams should still validate on their own tasks. But the trajectory is clear: open-weight models keep closing the capability gap while extending the cost advantage, and that combination is what ultimately shifts adoption.