Highest ARC-AGI-3 Scores, by Model (Standard Harness)
On ARC-AGI-3's Semi-Private set, run by ARC Prize on its provider-neutral Standard harness, GPT-6 Astra leads at 62.71%, ahead of GPT-6.1 Sol at 52.73% and Claude Opus 5 at 30.16%. Most of the other 16 models score under 5%.
ARC-AGI-3 Semi-Private, Standard harness %
GPT-6 Astra62.71
GPT-6.1 Sol52.73
Claude Opus 530.16
Gemini 3.8 Flash10.37
GPT-5.6 Sol7.78
GPT-6 Sol4.62
Grok 4.62.11
Grok 4.71.82
Measured by ARC Prize, each model's best reasoning-effort setting; leaderboard data generated Oct 5, 2026
Chart: the eight highest ARC-AGI-3 Semi-Private scores on ARC Prize's Standard harness, one bar per model at its best reasoning-effort setting.
As of (leaderboard data generation date, checked Oct 13, 2026)All scores measured by ARC PrizeTies share a rank
Highest ARC-AGI-3 Semi-Private scores by model on ARC Prize's Standard harness, as of 2026-10-05
Grok 4.20 Beta1xAI · listed as “Grok 4.20 (Beta Reasoning)”
xAI
Reasoning
0.09
3,775
—
Measured
Metric: the ARC-AGI-3 Semi-Private score ARC Prize publishes on its Verified leaderboard for the Standard harness, which gives every model the same provider-neutral interface and lets it carry forward notes it chooses to keep.2,13 ARC Prize defines 100% as an agent beating every game as efficiently as humans.15Selection: every model with a Standard-harness ARC-AGI-3 result in the leaderboard data (19 models), one row each at its highest-scoring reasoning-effort setting; ARC Prize tested each one itself, so every score is labeled measured and listed as a verified score.1,3 Community submissions, which ARC Prize does not verify by default, sit on a separate leaderboard and are not included.13Provider Adapter scores (a harness that keeps the provider's own reasoning state between requests and compacts long conversations) answer a different question, so they're shown for reference only, at their own best effort setting, and never ranked.12,13Run cost is ARC Prize's figure for the whole Semi-Private run at retail API prices, not a cost per task, rounded here to the dollar; ARC Prize's site code labels the ARC-AGI-3 figure "total cost, not per task".18,13 The testing policy states a $10,000 cap per Semi-Private run, yet several ARC-AGI-3 run costs exceed it and three are listed at exactly $10,000; costs are shown as published and nothing is ranked by them.13,1Left out: models on the leaderboard with ARC-AGI-1 and ARC-AGI-2 results but no ARC-AGI-3 result yet, including Claude Opus 5.5, Claude Fable 5 and 5.1, DeepSeek V4, Kimi K3, Qwen3.8-27B and GLM-5.3-Flash; and Kaggle competition entries, which run offline under separate rules.5,16Exposure: ARC Prize says the Semi-Private set is sent to model providers' APIs under zero-data-retention agreements, so it "acknowledge[s] the possibility of limited leakage over time".13 Figure labels: how they work. Results will be re-checked after ARC Prize 2026 announces its results on Dec 4, 2026.16
What ARC-AGI-3 measures
ARC-AGI-3 is the third version of the ARC-AGI benchmark series and its first interactive one. ARC Prize released it on March 25, 2026 as "hundreds of original turn-based environments" with "no instructions, no rules, and no stated goals": an agent has to explore each game, work out how it works and what winning looks like, and carry that forward across harder levels.14 ARC Prize says it tests exploration, modeling, goal-setting, and planning and execution, and that humans can solve 100% of the environments.12 Scoring rewards action efficiency, not just completion: a 100% score means beating every game as efficiently as humans.15 At launch, ARC Prize put the best frontier AI score at 0.51%.14
Video: ARC Prize's official ARC-AGI-3 launch video, showing the interactive game environments (ARC Prize channel).
The harness matters as much as the model. On the Standard harness GPT-6 Astra's best score is 62.71%; on the Provider Adapter harness it is 99.95%.1 ARC Prize calls the Standard harness "a consistent, apples-to-apples comparison across providers" and reports the two separately, which is why only Standard scores are ranked here.12 ARC Prize defines AGI as "a system's ability to acquire any skill a human can, as efficiently as a human can", and wrote after Astra's result that "we are not claiming that it is AGI" (September 3, 2026).12
The older ARC-AGI-2 has bunched up near the top: on its Semi-Private set GPT-6 Astra (Max) scores 95.0%, GPT-6.1 Sol (Max) 94.17% and Claude Opus 5.5 (High) 93.33%, against a human panel listed at 100%.17 That leaves ARC-AGI-3 as the version that still separates frontier models.
The top 10
#1
GPT-6 Astra (OpenAI)
62.71%
OpenAI's GPT-6 Astra (listed release September 2, 2026) scored 62.71% at Max reasoning effort for a run cost of $26,098, the highest Standard-harness result on the board.3 ARC Prize found that higher effort settings generally cost less for Astra because it solves games in fewer actions; at no reasoning effort it scored 35.2% for $49,791.12 On the Provider Adapter harness it reached 99.95% at High effort, and in that harness at Max it used fewer actions than the median tested human on 96.0% of levels.1,12
#2
GPT-6.1 Sol (OpenAI)
52.73%
GPT-6.1 Sol, released September 29, 2026, scored 52.73% on the Standard harness at Max effort, with a listed run cost of $7,632, against Astra's $26,098.4,1 On the Provider Adapter harness it reached 96.4% at XHigh effort. The same model posts 94.2% on ARC-AGI-2 Semi-Private at Max effort, which shows how much harder the interactive version still is.4
#3
Claude Opus 5 (Anthropic)
30.16%
Anthropic's Claude Opus 5 scored 30.16% at High effort, for $20,657, and was the highest-scoring model on ARC-AGI-3 as of July 24, 2026, according to ARC Prize.5,1 ARC Prize says it completed five Public Demo environments that no model had beaten before. Because of a short testing window, ARC-AGI-3 was run only at High effort, and it has no Provider Adapter result.5
#4
Gemini 3.8 Flash (Google)
10.37%
Google's Gemini 3.8 Flash, released September 2, 2026, is the highest-scoring model from outside OpenAI and Anthropic: 10.37% on the Standard harness at High effort, with a $4,398 run cost.6,1 On the Provider Adapter harness, also at High effort, it reaches 35.0%. On ARC-AGI-2 Semi-Private it scores 89.2%.6
#5
GPT-5.6 Sol (OpenAI)
7.78%
GPT-5.6 Sol (July 9, 2026) scored 7.78% on Semi-Private at Max effort, for $25,064, and averaged 13.33% on the Public set.7,1 ARC Prize says it was the first model to win an ARC-AGI-3 public game (ft09, at 87%), and attributes its edge to orienting itself correctly in a new environment before acting.7
#6
GPT-6 Sol (OpenAI)
4.62%
GPT-6 Sol, released September 22, 2026, scored 4.62% on the Standard harness at Max effort for $5,554.8,1 Its Provider Adapter result at the same Max setting is 23.03%. It scores 89.6% on ARC-AGI-2 Semi-Private.8
#7
Grok 4.6 (xAI)
2.11%
xAI's Grok 4.6 was tested on ARC-AGI-3 only at XHigh effort, scoring 2.11% "at a total cost of $5,612", in ARC Prize's words.9 At the same setting it scores 67.1% on ARC-AGI-2 Semi-Private.9
#8
Grok 4.7 (xAI)
1.82%
Grok 4.7, released September 21, 2026, scored 1.82% on the Standard harness at XHigh effort, slightly below its predecessor, with a $2,732 run cost.10,1 Its best Provider Adapter score is 10.0%, at High effort.10
#9
Claude Opus 4.8 (Anthropic)
1.52%
Claude Opus 4.8, listed with a June 1, 2026 release date, scored 1.52% at High effort. Its run cost is listed at exactly $10,000, the per-run cap in ARC Prize's testing policy.1,13 Its successor, Claude Opus 5, scored 30.16% on the same harness.1
#10
GPT-5.6 Terra (OpenAI)
0.80%
GPT-5.6 Terra, the middle tier of the GPT-5.6 family, scored 0.80% on Semi-Private at Max effort and averaged 2.3% on the Public set.11 ARC Prize notes that its XHigh setting scored below Max, "signaling that its ceiling is a comprehension/planning limitation rather than a budget one".11