ARC-AGI-3
A bird's-eye view of the benchmark: what it measures and every AI IQ chart built on it.
| # | Model | Effort | Score | Cost (V3) | Completion time (est.) | Weights |
|---|
About ARC-AGI-3
What ARC-AGI-3 measures
ARC-AGI-3 measures abstract reasoning in interactive game-like environments. An agent must explore, learn from feedback, and adapt its strategy to complete each level. Scores report Relative Human Action Efficiency, while Cost (V3) records the evaluation cost published by the ARC Prize leaderboard.
Reading the cost view
Each line is one model run at several reasoning-effort levels on ARC-AGI-3. Every point shows the score and Cost (V3) at that effort, with cheaper runs on the left. Up and to the left is better; a line that stays flat while moving right is buying little extra score with the extra spend. Generation is judged among models with a published result on this benchmark, so the newest scored model in each lab's line counts as current even when a newer model has no result yet; use the Generation filter to include previous and older models.
Reading the completion-time view
Each point is a public model's ARC-AGI-3 score against its estimated completion time (time to last token) for the input and output lengths set by the sliders. Timing comes from the site's per-model speed measurements, not from the benchmark run, so every point is an estimate and labels carry a ~ prefix. The newest generation with a published result is shown by default. Up and to the left is better.
Reading the score bars
Bars rank models by the source-backed benchmark value used for this chart. Longer bars indicate higher published scores.
Reading the model table
Each row is one canonical model: the score shown is the same best published configuration the score bars use. Models with several published reasoning-effort levels expand to one sub-row per level, with the score and Cost (V3) the source reports for that level. Completion time is the site's estimated time to last token for a 10K-token input and 1K-token output (marked ~), not a benchmark run time. The top 20 rows are shown by default; click a column heading to sort, and missing values always sort last.