Project benchmark

ARC-AGI-2 · Abstract reasoning

ARC Prize · 271 configurations with published scores. Review each metric and its test conditions separately.

Report identity & conditions

Data as of2026-10-19· Report published—. See individual records for run dates and test conditions.

Leaderboard updated2026-10-15T13:43:29.214Z

Run configurations remain separate, including different reasoning settings.
Individual test dates are not given. Update dates and model release dates from the report remain separate and do not substitute for actual test dates.
Costs follow the report’s run scope and dollar amounts; they are not model API unit prices.
ARC-AGI-2 · Semi-private evaluation set. Benchmark versions are compared separately.
Task score on the semi-private evaluation set, as a percentage. It is not ranked across benchmark versions or evaluation scopes.
Cost per task as reported, in USD/task. Total costs and task counts are not invented. This is neither cost per successful task nor a current API quote.
Human reference results are shown separately and excluded from model/system rankings and cost comparisons.

Left metricRight metric

Compare metrics within the same report. Each chart explains its direction and scope.
See the details for full values and run conditions.

Task score

Task score on the semi-private evaluation set, as a percentage. It is not ranked across benchmark versions or evaluation scopes.: showing 8 of 271 records with values. See details for configurations and dates.

Chart data · 8 entries

Name	%
GPT-6 Astra (Max)	95
GPT-6.1 Sol (Max)	94.167
Claude Opus 5.5 (High)	93.333
GPT-6 Astra (XHigh)	93.333
Claude Opus 5.5 (XHigh)	92.5
GPT-5.6 Sol (Max)	92.5
GPT-6 Astra (High)	92.083
GPT-6 Astra (Medium)	92.083

Cost per task

Cost per task as reported, in USD/task. Total costs and task counts are not invented. This is neither cost per successful task nor a current API quote.: showing 8 of 271 records with values. See details for configurations and dates.

Chart data · 8 entries

Name	USD / Task
GPT-5 Nano (Minimal)	0.003
GPT-5 Nano (Low)	0.003
GPT-4.1-Nano	0.004
GPT-6 Luna (None)	0.004
DeepSeek V4 Flash 0731 (None)	0.004
Gemini 2.0 Flash	0.004
Qwen3-235b-a22b Instruct (25/07)	0.004
GPT-6 Luna (Low)	0.005

Human reference

Human baselines are shown separately and are not ranked alongside models and systems.

Human Panel

Task score: 100%
Cost per task: 17 USD / Task

Full results & model versions

271 / 271configurations have published scores in this scope. Scored configurations are listed; missing values remain “Not provided.” Different run configurations of the same model are listed separately.