ARC-AGI-2 Leaderboard: What GPT-6 Astra's 95 Percent Score Actually Means (2026)

As of September 18, 2026, GPT-6 Astra sits at the top of the single model ARC-AGI-2 leaderboard at 95 percent, ahead of GPT-5.6 Sol at 92.5 percent and Claude Opus 5 at 90.4 percent. Average individual human performance on the same task set is 66 percent, so the frontier is now comfortably past what an unassisted human scores on the benchmark that was explicitly designed to resist the kind of memorization and pattern matching that made its predecessor easy to game. That headline invites a specific kind of coverage, models beating humans at a benchmark named after general intelligence, and it is worth being precise about what the number does and does not support before drawing that conclusion.

What ARC-AGI-2 is actually testing

ARC-AGI-2 is a set of abstract visual reasoning puzzles: given a small number of example input and output grids that share an unstated transformation rule, produce the correct output grid for a new input the system has not seen. The task is deliberately resistant to the two easiest ways a benchmark gets gamed. It is not language, so there is no shortcut through memorized text patterns, and each puzzle instance is constructed to be novel rather than a variation on a publicly known type, so a system cannot simply recall a similar solved example from training data. That design is exactly why ARC-AGI-2 has held up as a harder benchmark than most coding and knowledge tests, and why 23 models have now been evaluated on it without the near universal saturation you see on sets like SWE-bench Verified, which we took apart in what the SWE-bench Verified score actually measures.

But resistant to memorization is not the same as a measurement of general intelligence, and the benchmark's own name does more work in press coverage than the task design supports. ARC-AGI-2 measures the ability to infer a transformation rule from a small number of visual examples and apply it correctly to a new case. That is a real and useful capability, closely related to few shot pattern induction, and it correlates with reasoning quality in ways that are worth tracking. It is not a test of planning across a long task, tool use, memory across a session, or the kind of grounded interaction with a real environment that terms like general intelligence typically imply. A model can score well here and still fail badly on an open ended agentic task, and a model can be a weaker fit for ARC-AGI-2's specific visual puzzle format while being the better choice for your actual workload.

The scaffolded systems at the top are the most instructive part

The two highest scores on the leaderboard, Confluence Lab at 97.9 percent and Imbue's Darwinian Evolver at 95.1 percent, are not single models. They are scaffolded systems: a base model wrapped in additional structure such as multiple sampling attempts, self verification passes, or search over candidate solutions before committing to a final answer. That these systems outscore every raw single model, including the leaderboard leading GPT-6 Astra, is the clearest evidence on the page that the score is measuring the combination of a model and the scaffold around it, not model capability in isolation.

This is the same lesson we drew out from a different angle in GPT-6 Astra versus GPT-5.6 Sol on whether the upgrade is worth it: a benchmark score attributed to a model name is frequently a score attributed to a specific configuration of that model, with a specific scaffold, prompt strategy, and compute budget behind it. On ARC-AGI-2 specifically, the gap between a raw model and a scaffolded system built on comparable underlying capability is close to three full points, larger than the gap separating the top three raw models from each other. If you are evaluating a model based on a headline ARC-AGI-2 number, the first question to ask is whether that number came from the raw model or from a scaffold you would also have to build and maintain to reproduce it.

Why beating the human baseline is not the milestone it sounds like

Average individual human performance on ARC-AGI-2 is 66 percent, and frontier systems clearing 90 percent is a genuine result worth noting. But the human baseline here is a specific, narrow measurement: individual humans, working alone, on a puzzle format most people have never practiced, under whatever time constraint the evaluation used. It is not a measure of peak human reasoning ability, collaborative human problem solving, or human performance with tools and iteration available, all of which are the more relevant comparisons if the question is what these systems can do relative to human capability in a real work context. Treat "above the human baseline" as a statement about this specific benchmark's specific baseline, not a general claim about reasoning superiority.

What to actually take from the leaderboard

Track the trend, not the single number. ARC-AGI-2 scores have moved substantially within 2026, and a benchmark that has not saturated is genuinely informative about relative model improvement over time in a way that a saturated benchmark like SWE-bench Verified no longer is. Separate raw model scores from scaffolded system scores when you compare, since the two are measuring different things wearing the same leaderboard. And treat a strong ARC-AGI-2 score as evidence of few shot pattern induction ability specifically, which is a real and transferable skill for tasks that resemble it, rather than as evidence for capability on tasks that do not, such as long horizon agentic work or grounded tool use. We covered how that category level nuance plays out for coding agents specifically in the coding agent adoption and Terminal-Bench landscape.

FAQ

What does ARC-AGI-2 actually measure? The ability to infer a transformation rule from a small number of example visual grids and apply it correctly to a new, unseen case. It is designed to resist memorization and pattern recall from training data.

Does a high ARC-AGI-2 score mean a model has general intelligence? No. It means the model or system is strong at few shot abstract pattern induction on this specific visual puzzle format. That capability does not transfer automatically to planning, tool use, or long horizon agentic tasks.

Why do scaffolded systems score higher than raw models? Systems like Confluence Lab and Imbue's Darwinian Evolver add structure such as multiple sampling attempts or self verification on top of a base model, which the leaderboard credits as a single score. The result reflects the model and the scaffold together, not the model alone.

Is beating the 66 percent human baseline a major milestone? It reflects average individual human performance on this specific benchmark under its specific conditions, not peak human reasoning or human performance with tools and iteration available. Read it as one data point, not a general claim about reasoning superiority.

When this applies to your stack

If you are using ARC-AGI-2 or any single benchmark score to choose a model, separate the raw model result from any scaffold used to produce it, and weight the benchmark by how closely its task format resembles the work you actually need done. A strong pattern induction score is a real signal for tasks that look like pattern induction and a weak signal for everything else. Building the evaluation harness that tells the difference for your specific use case is the kind of applied AI work we do at Contra Collective.