On July 24, Anthropic’s Claude Opus 5 cleared the highest score ever recorded on ARC-AGI-3 — the benchmark designed to be the hardest general-reasoning test in AI — and while doing it, wrote an algebraic equation that no frontier model had ever produced, according to ARC Prize’s independently administered results. The equation, generated spontaneously at action 23 of one environment, read: 4_center = 2×axis − 5_center. ARC Prize, which administered the test independently, called it the first explicit reflection equation from any model in their analysis, as documented in their official July 24 announcement.

Opus 5 scored 30.2% — nearly quadrupling GPT-5.6 Sol’s previous record of 7.8%, per the ARC Prize results page. It solved five environments no AI had previously beaten, four of them at or above human efficiency, and pushed the total number of solved public demo environments from one to six, as confirmed in ARC Prize’s July 24 X post. The gap to its nearest competitor is not incremental. Anthropic’s own prior generation — Fable-class models — reached approximately 20%, according to ARC Prize. GPT-5.6 Sol, the previous record holder, sits at 7.8% at Max reasoning effort.

The score alone would have moved the story. The method behind it changed what the story was about.

ARC-AGI-3: What the Benchmark Actually Tests

ARC-AGI-3 is deliberately unlike every other AI evaluation in regular use. Where most benchmarks ask a model to retrieve knowledge or complete a task it has seen some version of before, ARC-AGI-3 drops an AI agent into a novel interactive environment with no instructions, no language, and no domain knowledge that applies — and measures how efficiently it figures the environment out, as detailed in the ARC-AGI-3 technical paper.

The benchmark uses only what its designers call Core Knowledge priors: the basic object-recognition and physical-intuition capabilities that human infants develop before any formal education, as explained in the ARC Prize Foundation technical report. Agents must explore to learn each environment’s rules, then execute to reach the goal. When ARC-AGI-3 launched in March 2026 at Y Combinator in San Francisco, the results were stark: human testers cleared 100% of environments, while the best frontier AI model managed 0.37%, per the technical paper’s launch data.

The benchmark differs from its predecessors in a structurally important way. ARC-AGI-1 and ARC-AGI-2 presented static grid puzzles — given a few examples of an input-output transformation, infer the rule and apply it. ARC-AGI-3 replaces the static puzzle with a turn-based interactive game. An agent must take actions, observe results, update its beliefs about what the environment is doing, and form an efficient strategy — all without being told what the game is or what winning looks like, as the technical paper’s design section explains.

Why Scoring Rewards Thinking, Not Guessing

Performance is measured by a metric called RHAE — Relative Human Action Efficiency — pronounced “Ray.” For each completed environment, an AI’s score is calculated as the square of the ratio of the human baseline action count to the AI’s action count, per the ARC Prize technical report formula. If human testers completed an environment in 10 actions and an AI took 100, the AI scores (10/100)² = 1%. Using 30 actions would yield (10/30)² = 11.1%. Using 20 would yield 25%.

The squared term is what makes ARC-AGI-3 resistant to brute-force approaches. Doubling the number of actions does not halve the score — it quarters it. An agent that tries everything will score near zero regardless of whether it eventually succeeds. Only an agent that quickly forms an accurate model of how the environment works, then acts on that model efficiently, will produce a meaningful RHAE score. Levels are capped at five times the human action count; anything beyond that earns zero for that level, though scores above 100% are possible for agents that outperform the second-best human on a given environment, as specified in the technical report’s scoring rules.

This design directly penalizes what prior-generation AI systems did best: exhaustive pattern-matching. It rewards what they have historically done worst: rapid hypothesis formation, belief updating under uncertainty, and efficient execution of a learned strategy — what psychologists call fluid intelligence.

Opus 5’s Score in Context

Four months after launch, where every frontier AI had scored below 0.37%, Claude Opus 5 scored 30.2%, according to ARC Prize’s verified results. This matters not just because the number is large but because of who measured it. ARC Prize independently administered the run — running the model itself against the test environments, recording full reasoning traces, and publishing results without Anthropic controlling the harness, configuration, or output, as explained by Digital Applied’s independent analysis.

Almost every other benchmark number in any frontier AI launch announcement is vendor-run: the lab selects the benchmark version, configures the evaluation harness, picks the reasoning effort level, and reports the result, per Digital Applied’s buyer’s guide. ARC Prize’s independent administration means the 30.2% figure has been verified by a third party with no stake in the outcome. Full replay logs, including Opus 5’s complete reasoning traces for each environment, are publicly downloadable from the Opus 5 scorecard page.

The five newly cleared environments are designated ar25, ft09, lp85, r11l, and s5i5. Of these, four were completed at or above human efficiency levels — meaning Opus 5 solved them with fewer or comparable actions to the best human testers. The sixth solved environment (vc33, at 98.8% RHAE) was already in the record before Opus 5’s run, as confirmed in the Opus 5 environment scorecard.

Nineteen of the 25 public demo environments remain unsolved by any AI.

The Equation That Changed the Interpretation

The 30.2% score earned its headline. The behavior that produced it earned the researchers’ attention.

ARC Prize’s analysis team noted that during Opus 5’s run through one environment, the model generated a formal algebraic representation of a spatial relationship without being prompted to do so, as documented in the ARC Prize official X post. At action 23 of that environment, the model’s reasoning trace contained: 4_center = 2×axis − 5_center. This is a reflection equation — it encodes the geometric relationship between an object’s center, a symmetry axis, and its reflected position, as ARC Prize explained in their July 24 announcement.

ARC Prize described it as the first explicit reflection equation produced by a frontier model in their analysis. The equation is not a prose description of reflection, and it is not a retrieved example from training data about reflection — it is a formal symbolic statement constructed by the model to represent a spatial relationship it needed to navigate to complete the environment.

Why does this matter technically? ARC-AGI-3 environments contain no language instructions and no mathematical prompts. The model encountered an environment involving spatial symmetry, represented what it observed as a formal equation, and then acted on that representation. This suggests Opus 5 is building intermediate symbolic representations of environment state — essentially constructing a private formal language to reason about what it sees before deciding what to do next. ARC Prize attributed the lead on ARC-AGI-3 directly to this capacity, noting that stronger logical reasoning enables more autonomous exploration, planning, and execution across unfamiliar environments.

What Independent Puzzle Testing Found

Researchers outside ARC Prize ran Opus 5 through a separate evaluation that points toward a more complex picture.

Guanghan Ning, whose Witness benchmark tests models on held-out interactive puzzle games built in the same spirit as ARC-AGI-3, found that Opus 5’s gains do not transfer proportionally to different puzzles, as reported in his July 24 X post. On Witness composite scores, using the same harness and token budget across all models, Opus 5 scored 43.4 ± 3.2 — a statistical tie with Kimi K3 (42.8 ± 1.9) and Fable 5 (43.8 ± 9.7), per Ning’s Witness data. It improved over Opus 4.8’s Witness score of 34.8, but the improvement was far smaller than the near-4x jump on ARC-AGI-3.

Ning’s replay traces offered a specific explanation. On the Witness benchmark’s most classic puzzle — one built around familiar mechanics — Opus 5 stated the hidden rules before taking its first action and then played a near-optimal solution across all test seeds, as documented in Ning’s analysis. Zero exploration. The model apparently already knew this genre. On the most novel Witness puzzle — one built around unusual mechanic combinations that cannot be pattern-matched — Opus 5 scored below Opus 4.8. Ning interpreted this pattern as consistent with training on genre-specific puzzle data, though the Witness benchmark cannot identify what data Anthropic used.

Ning later clarified that Opus 5 did show gains across Witness as a whole, though far smaller than on ARC-AGI-3. He compared the pattern to the evolution of coding benchmarks: as a new benchmark becomes a major target, it receives training effort first; as researchers add more edge cases, models learn to handle a broader range of tasks, as reported by The Decoder.

Greg Kamradt, president of ARC Prize, said the Witness results do not rule out broader reasoning gains, according to The Decoder’s coverage. He noted that Opus 4.8 also outperformed Opus 5 on some individual ARC-AGI-3 environments despite trailing it by a wide overall margin — which means environment-level regressions are not incompatible with genuine aggregate reasoning improvements. Judging the breadth of the gain would require detailed results from more unfamiliar tasks.

What Anthropic Has Not Explained

There is also a methodological note in the comparison that matters for interpretation. GPT-5.6 Sol’s 7.8% was evaluated at Max reasoning effort — the highest available effort setting, per the ARC Prize results page comparison data. Opus 5’s 30.2% was evaluated only at High reasoning effort, because the pre-launch testing window was short, as ARC Prize confirmed on their results page. The gap is enormous even accounting for the effort mismatch — High and Max often produce comparable results on mature benchmarks — but it is not an effort-matched comparison.

Anthropic has not publicly explained what accounts for the ARC-AGI-3 gain. Targeted data labeling toward ARC-AGI-3’s puzzle formats, and reinforcement learning on interactive environment tasks, are plausible mechanisms — and Opus 5 was developed after ARC-AGI-3’s format became public, which may have allowed Anthropic to target the benchmark’s specific skills. Whether that targeting explains all, most, or some of the gain cannot be determined from the published materials, as The Decoder’s analysis concluded.

Does Better ARC-AGI-3 Mean Better AI?

The benchmark’s own designers are careful about what a high score implies. ARC-AGI-3 was built to measure one specific gap: the gap between how efficiently an AI and a human can learn to navigate a novel interactive environment, as described on the ARC Prize results page. A score of 100% would mean an AI agent can solve every environment as efficiently as the second-best human tester — the benchmark’s operators are explicit that this would not mean AGI had been achieved. It would mean one particular, measurable gap between AI and human adaptive learning had been closed.

By that measure, 30.2% represents significant but partial progress — particularly given that the Witness counter-evidence indicates the gains may be concentrated on environments similar to what Opus 5 trained toward. The algebraic notation behavior is harder to explain away: writing a formal symbolic equation for a spatial relationship encountered in a novel environment is not the same as retrieving a stored pattern. Whether it represents a genuine capability advance, an emergent behavior produced by scale, or a more specific learned behavior remains an open empirical question — and one ARC-AGI-3’s design cannot yet definitively answer.

This is the deepest implication of the Witness results: a benchmark designed to test the kind of reasoning that cannot be acquired from training data can itself become a training target, blurring the line between genuine fluid reasoning and very sophisticated crystallized-intelligence retrieval. That distinction — whether an AI is building new reasoning structures or retrieving learned ones — is the fundamental scientific question ARC-AGI-3 exists to probe. For now, 30.2% is the clearest signal the field has that the gap is narrowing. How much of the narrowing reflects genuine reasoning capability, and how much reflects Anthropic’s ability to target a specific benchmark, is the right question to carry forward.

Frequently Asked QuestionsWhat is ARC-AGI-3 and how does it score AI performance?

ARC-AGI-3 is an interactive benchmark developed by ARC Prize Foundation that tests whether AI agents can learn the rules of a novel environment, form a strategy, and complete it efficiently — without instructions, language, or prior knowledge of the task. It scores using a metric called RHAE (Relative Human Action Efficiency): for each completed environment, an AI’s score equals the square of the ratio of the human baseline action count to the AI’s action count, as defined in the ARC-AGI-3 technical paper. If human testers take 10 actions and an AI takes 100, the AI scores 1% for that level. The squared formula heavily penalizes slow, exploratory approaches and rewards rapid, accurate hypothesis formation.

How does Claude Opus 5’s 30.2% compare to other frontier models?

It is the highest recorded score on ARC-AGI-3. The previous record was 7.8%, set by GPT-5.6 Sol at Max reasoning effort. Anthropic’s own prior Fable-class models scored approximately 20%. Opus 5 reached 30.2% at High reasoning effort — the two are not effort-matched, but the gap remains large even accounting for this. On the older ARC-AGI-1, Opus 5 scored 97.5% (essentially at saturation). On ARC-AGI-2, it scored 90.4%, all as verified on the ARC Prize official results page.

What does the reflection equation actually tell us about Opus 5?

ARC Prize documented that Opus 5, at action 23 of one environment, wrote 4_center = 2×axis − 5_center — a formal algebraic encoding of a spatial reflection relationship, as confirmed in ARC Prize’s official analysis. This is notable because the environment contained no mathematical prompts and no language instructions. The model appears to have constructed a symbolic representation of the spatial structure it was navigating, which ARC Prize described as the first explicit reflection equation produced by any frontier model in their analysis. Whether this represents a genuine emergent symbolic reasoning capability or a learned behavior triggered by familiar spatial patterns is the question the Witness benchmark’s counter-evidence puts squarely on the table.

Can ARC-AGI-3 reliably distinguish genuine new reasoning from sophisticated pattern retrieval?

This is the central open question, and the Witness benchmark results make it concrete. Guanghan Ning’s Witness evaluation — using held-out puzzle games built around similar principles — found Opus 5 performing at or above human efficiency on a familiar puzzle genre before taking a single exploratory action, then scoring below its predecessor on a genuinely novel mechanic combination. A benchmark designed to measure the kind of reasoning that cannot be learned may itself become a training target. ARC Prize’s own 2025 technical report explicitly flagged that “current frontier AI reasoning performance remains fundamentally constrained to knowledge coverage, giving rise to new forms of benchmark contamination,” as documented in arXiv:2601.10904. The 30.2% score is the best signal the field currently has that the human-AI gap is narrowing — interpreting exactly what it measures requires holding both the score and the Witness results in view.