DeepSeek V4 Flash 0731 hits 89% on ARC-AGI-1 at two cents a task

DeepSeek has posted a new reasoning model, V4 Flash 0731, and its ARC-AGI results are the cheapest high score the leaderboard has seen. On the ARC Prize results page, the model scores 89.0% on ARC-AGI-1 Semi-Private at max effort, priced at $0.02 per task. The same run reaches 61.4% on ARC-AGI-2 Semi-Private at $0.04 per task.

The headline number is the ARC-AGI-1 result. A year ago, 85% on that benchmark was frontier territory. DeepSeek is now selling it at the cost of a rounding error in most inference budgets. The model ships with three reasoning variants, labeled Max, High, and Low, and the scores scale predictably: 89.0%, 87.0%, and 84.0% on ARC-AGI-1, and 61.4%, 56.0%, and 46.0% on ARC-AGI-2. No ARC-AGI-3 score is listed.

The more interesting number is the drop. The gap between ARC-AGI-1 and ARC-AGI-2 at max effort is 27.6 points. That is not a small degradation. It is the shape of a model that has learned the first benchmark’s tricks well and still struggles when the second benchmark changes the rules.

What the task-level data shows

The ARC Prize page includes pass/fail tables for every task in the public evals. The ARC-AGI-2 Public Eval has 120 tasks, and the pattern is revealing. Many tasks that the Max variant solves, the Low variant fails. Task 0934a4d8 is a clean example: Max fails it, High fails it, and Low fails it. Task 135a2760 passes at all three levels. Task 13e47133 fails at all three.

The monotonicity is mostly clean. When Max passes and Low fails, the High variant usually sits in between. But there are exceptions worth noting. Task 4a21e3da shows Max passing, High failing, and Low passing, a non-monotonic result that suggests some tasks are solved by luck of the sampling rather than by a reliable reasoning mechanism. Task 7b5033c1 shows Max failing, High failing, and Low passing, which is harder to explain and points at instability in the search process.

On ARC-AGI-1 Public Eval, the 400-task table is mostly green. The model passes at all three levels on the vast majority of tasks. A handful of tasks fail everywhere, including 0d87d2a6, 1acc24af, 212895b5, 25094a63, 73ccf9c2, 8b28cd80, 8fbca751, d017b73f, da515329, and f3b10344. Ten tasks out of 400 is 2.5%, which matches the 84% Low score on that benchmark only if the Low variant also loses some tasks that Max wins. The task-level data confirms that: 14754a24, 16b78196, 477d2879, and 79fb03f4 all show Max passing while Low fails.

The ARC-AGI-2 wall is the real story

ARC-AGI-2 was designed to be harder than its predecessor. The benchmark’s authors deliberately removed the pattern-matching shortcuts that models exploited on ARC-AGI-1, and the new tasks require compositional reasoning and multi-step abstraction. DeepSeek’s 61.4% is respectable, but it is not close to the near-saturation the model achieves on ARC-AGI-1.

The interesting question is what the 27.6-point gap costs. At $0.04 per task, running the full ARC-AGI-2 Semi-Private set is cheap. The cost is not the blocker. The blocker is that the model’s reasoning does not transfer cleanly from one benchmark generation to the next. This is the same pattern the field has seen before: models overfit the distribution of their training and evaluation data, and each new benchmark generation exposes the brittleness.

DeepSeek’s paper and model card, linked from the results page, presumably describe the architecture and training approach. What the ARC Prize page itself makes clear is that this is a “Flash” model, positioned as a fast, cheap inference option rather than a max-capability flagship. The pricing confirms the positioning: $0.02 per task on ARC-AGI-1 is an order of magnitude cheaper than what frontier labs charged for similar scores a year ago.

What this means for the inference economy

The two-cents-per-task price is the economic signal. It tells you that DeepSeek has found a way to compress the reasoning compute needed for these tasks, or that the model is efficient enough that the marginal cost of a task is nearly nothing. Either way, the price pressure on the rest of the market is real. OpenAI, Anthropic, and Google have all priced reasoning models at a premium. A model that hits 89% on ARC-AGI-1 at $0.02 per task makes that premium harder to justify for any customer whose workload looks like ARC-AGI-1.

The ARC-AGI-2 gap is the counterweight. If your workload looks like ARC-AGI-2, the cheap model is not good enough, and you still need the expensive frontier models. The market is bifurcating: cheap models for the easy distribution, expensive models for the hard distribution. DeepSeek is competing hard on the first axis and has not yet closed the gap on the second.

The open-source angle

DeepSeek has historically released its models as open weights, and the results page links to a model page and paper. If V4 Flash 0731 follows the pattern, the weights will be downloadable and the community will be able to fine-tune and distill it. That changes the calculus for everyone building on top of these models. A two-cent-per-task open-weights model that scores 89% on ARC-AGI-1 is a platform shift for the agent-building community.

The task-level data on the ARC Prize page is a gift to researchers. Anyone can now see exactly which tasks the model fails at all three reasoning levels, and those failure cases are the raw material for the next round of training data. The 13e47133 failures and the 4a21e3da non-monotonic results are the kind of concrete evidence that drives the next architecture change.

The honest caveat

ARC-AGI scores are not the same as general intelligence. The benchmark is narrow, and a model can score well on it by learning the specific inductive biases the benchmark rewards. The 89.0% on ARC-AGI-1 is impressive, but the 61.4% on ARC-AGI-2 is the more honest measure of how far the model is from robust generalization. The gap between the two is the distance between a model that has mastered a known distribution and a model that can handle the unknown.

DeepSeek has published the scores, the pricing, and the per-task breakdown. That transparency is good for the field. The next step is for the research community to dig into the failure cases and figure out why the model drops 27.6 points when the benchmark gets harder. The answer to that question, more than the 89.0% headline, is what will determine whether the next benchmark generation closes the gap or widens it.

The per-task table for ARC-AGI-2 shows 120 tasks, and the model fails all three variants on a substantial minority of them. Those tasks are the frontier now. Whoever solves them first, at DeepSeek’s price point, will own the next round of the reasoning race.