Claude Opus 5 Hits 30% on ARC-AGI-3, a New Interactive Benchmark Where Rivals Stall Below 10%

Anthropic's Claude Opus 5 has become the first model to clear 30 percent on ARC-AGI-3, the newest iteration of the abstract reasoning benchmark that tests AI agents inside interactive environments rather than handing them static puzzles. The score, 30.2 percent at High reasoning effort, is roughly three times the next-best competitor and twenty times what the previous generation managed 1.

A 30 percent pass rate on a reasoning benchmark is not a victory lap. It is a diagnostic. Claude Opus 5 fails on nearly 70 percent of the novel interactive tasks the benchmark presents, including three of the 25 Public Demo environments where it scores zero 1. That gap is what the score actually measures.

The numbers behind the gap

ARC-AGI-3's Public Demo suite contains 25 environments, each requiring an agent to reason through a scenario it has not encountered in training 1. Claude Opus 5 (High) solved enough of them to post 30.16 percent. The next-best model, GPT-5.6 Sol, scored 7.8 percent. Claude Opus 4.8, which launched weeks earlier, managed 1.5 percent 2. Anthropic describes the result as roughly three times the next-best model's score 2. Against the previous generation, the jump from 1.5 to 30.2 is twentyfold 3.

The per-environment breakdown reveals a model that improved unevenly. Claude Opus 5 completed five environments at 100 percent that no prior model had solved 1. It also scores zero on BP35, G50T, and SC25. This is not uniform progress across the board. It is a model that cracked certain categories of novel reasoning entirely while remaining unable to make progress on others.

What ARC-AGI-3 actually tests

The benchmark's own terminology signals the shift. ARC-AGI-1 and ARC-AGI-2 present models with tasks: static puzzles with a fixed input and a correct output. ARC-AGI-3 presents environments where agents must act and observe consequences 1. Claude Opus 5 scores 97.5 percent on ARC-AGI-1 at Max reasoning effort, a near-saturation result. Its 30.2 percent on ARC-AGI-3 maps the distance between recognizing a pattern and navigating a situation you have never seen 1.

AI Tools Review describes ARC-AGI-3 as measuring whether a model can solve genuinely novel problems it has not seen patterns for, calling Opus 5's result "the clearest evidence that the gains are in reasoning rather than memorised benchmark structure" 3.

One caveat: due to the short testing window, ARC-AGI-3 was evaluated only at High reasoning effort, not at Max 1. The 30.2 percent may not represent Opus 5's ceiling.

The behavior behind the score

Anthropic reports that Opus 5 is substantially stronger at checking its own work and iterating until it succeeds, rather than stopping at a plausible first answer 3. That habit matters disproportionately in interactive environments, where an agent can test a hypothesis, watch it fail, and revise.

What 30 percent means

For builders deploying agents in production, a 30 percent success rate on novel interactive tasks means the error-handling and fallback architecture is carrying most of the load. For investors, the ARC-AGI-3 leaderboard is now the most honest map of the frontier because it ranks models by what they cannot do. The shift from tasks to environments is the structural change underneath this score. Static benchmarks measured how well models recognize patterns they have been trained on. ARC-AGI-3 measures how well they reason when the pattern does not exist yet. Claude Opus 5 is the first model to suggest that transition is tractable. It is also proof that tractable is not the same as solved.