Gemini 4 Argon Scores 97.5% on ARC-AGI-3, Approaching Human-Level
Key Highlights
ARC Prize published Gemini 4 Argon's results on ARC-AGI (Verified), and they are striking: on ARC-AGI-3, the benchmark widely regarded as the hardest test of genuine generalization, it scored 97.5% via a provider adapter harness at a cost of only 4,800 dollars. Google DeepMind thereby became the second frontier lab to break 95% on ARC-AGI-3. Many read this score as a landmark moment of AI approaching human-level abstract reasoning.
What It Does
ARC-AGI problems show a few example figure transformations and ask you to infer the rule and apply it to a new figure, testing true few-shot abstraction rather than rote memorization. Gemini 4 Argon's near-perfect ARC-AGI-3 score signals a qualitative leap in low-sample abstract reasoning; it also posted 91.3% and 98.5% on ARC-AGI-2 and ARC-AGI-1 respectively, at extremely low per-task cost of 0.16 to 0.32 dollars.
Technical Details
A telling detail is the harness gap: the provider adapter harness reached 97.5% while the standard harness managed only 49.8%, suggesting the score depends heavily on how the task is fed to the model. ARC-AGI-3 is deliberately built to defeat memorization, so high scores require real generalization, but also careful reasoning interfaces. The 4,800-dollar cost is already a steep drop from traditional approaches that needed massive sampling.
Versus Alternatives
The club of models past 95% on ARC-AGI-3 was tiny before this. Gemini 4 Argon's entry shows Google catching up on the hardest true-reasoning track. Versus rivals like OpenAI, DeepMind's edge lies in data and training methodology, but the standard harness's 49.8% is a reminder that stable high scores under any prompt are not yet here and that engineering contributes a large share of the headline number.
Industry Impact and Use Cases
ARC-AGI is considered one of the touchstones on the path to AGI. A model approaching humans here means AI is improving at tasks requiring abstraction from few examples, novel-domain planning, and rare-fault diagnosis. For enterprises, it hints at AI that delivers sound solutions from fewer examples and fuzzier requirements. Of course, the harness gap also shows evaluation methods need standardization, or scores will keep being inflated by engineering.
The lesson for the field is that ARC-AGI-3 is becoming the new bar for claiming genuine reasoning, and the harness gap is a caution that published numbers must be read with care. A 97.5% under one harness and 49.8% under another is not a contradiction so much as a reminder that how you ask matters as much as what the model knows. For practitioners, the takeaway is that frontier reasoning is now attainable but brittle across interfaces, and the cost of 4,800 dollars, while down, is still non-trivial for routine use. The practical guidance is to wait for standardized harness results before treating ARC-AGI-3 as solved, and to remember that the easier ARC-AGI-1 and ARC-AGI-2 scores, already near saturation, are no longer where the action is.
Looking ahead, the arrival of multiple labs past 95% on ARC-AGI-3 suggests the benchmark's role as a differentiator is ending, and the field will need harder tests of genuine reasoning and adaptation. The harness sensitivity already hints that we are measuring interfaces as much as intelligence, and the next chapter of evaluation must standardize how models are prompted and used. For practitioners, the practical implication is to stop treating any single ARC score as proof of AGI-adjacent capability and instead probe robustness across framings. The 4,800 dollar cost, while lower than brute-force sampling, still argues for reserving such models for tasks where the reasoning genuinely pays off rather than routine work a cheaper model handles.
The closing thought is that ARC-AGI scores, however impressive, are a snapshot of a moving target, and the field should resist declaring victory on any single benchmark. The harness gap is the real lesson, that how we evaluate is now as important as what we build, and standardized, transparent evaluation must keep pace with model capability. For buyers, the implication is to demand comparable, reproducible numbers rather than headline percentages. Gemini 4 Argon's result is genuinely important, but its lasting value will be in pushing the whole field toward harder, fairer tests of reasoning rather than in a single triumphant score.
The deeper lesson of the ARC-AGI-3 result is that reasoning benchmarks are evolving faster than the models, and that is healthy. ARC-AGI was designed to be hard for systems that merely memorize patterns, so a 97.5 percent score signals real progress in fluid, few-shot reasoning rather than overfitting to a known test set. The reported cost of roughly 4.8 thousand dollars per task is the part builders must weigh: at that price the model is a research instrument, not a production tool, and the cost curve will have to fall dramatically before such reasoning is economical at scale. The provider adapter harness and standard harness producing different numbers is also instructive, because it shows how much the scaffolding around a model affects the result. Teams should resist the temptation to read a single percentage as a verdict on capabilities. Instead, the useful takeaway is that the field now has a harder, fairer test, and that the best systems are closing the gap with human-level performance on tasks that were recently considered out of reach. Gemini 4 Argon sets a marker; the next releases will show whether it was a peak or a stepping stone.