Guide
What Is ARC-AGI-3?
ARC-AGI-3 is the ARC Prize Foundation's interactive reasoning benchmark. It drops an agent into a novel environment with no natural-language instructions and asks it to work out what matters, what the goal is, and how to get better at reaching it. That makes it one of the few public benchmarks that measures the thing an agent actually does, rather than the thing a chat model does.
The One-Sentence Version
ARC Prize describes it as a benchmark that “challenges AI agents to explore novel environments, acquire goals on the fly, build adaptable world models, and learn continuously.” The scoring anchor is human competence: a 100% score means agents beat every game as efficiently as humans do. Not “eventually solve it” — as efficiently.
How It Differs From ARC-AGI-1 and 2
The earlier ARC generations are static. You are shown a handful of input-output grid examples, you infer the transformation, you apply it to a new grid. It is a puzzle you look at, reason about once, and answer.
ARC-AGI-3 is interactive. There is no worked example to study and no prompt telling you the objective. The agent has to act, observe what changed, and revise its understanding — repeatedly, inside the environment. That shift matters because it tests the failure mode that actually breaks agents in production: not “can it reason about a described problem” but “can it figure out what is going on when nobody described anything.”
What It Measures
| Criterion | What it is testing |
|---|---|
| Human-solvable environments | Every environment can be beaten by a person, so a low score is a gap in the agent rather than an impossible task. |
| Skill acquisition over time | The agent is measured on whether it gets better inside an environment, not only on whether it eventually wins. |
| Long-horizon planning with sparse feedback | Rewards are thin and delayed, so an agent cannot grope its way forward one step at a time. |
| Experience-driven adaptation | Strategy has to change in response to what the environment does, not follow a plan fixed at the start. |
Reading the Scores Honestly
Two cautions before you use an ARC-AGI-3 number to pick a model.
- The harness is part of the score. The same model evaluated through different harnesses can post very different results, because the harness decides how the model perceives the environment and how its actions are executed. A headline number without its harness is close to meaningless. When comparing two models, check they were run the same way.
- Cost per task is part of the result. Scores on this benchmark are reported alongside what it cost to achieve them, and the top of the leaderboard has historically been expensive. A model that scores highly by spending an enormous amount of compute per task is telling you something real about capability and nothing encouraging about your monthly bill.
Because the leaderboard moves and this page would go stale the moment it froze a figure, read the current standings at the ARC Prize leaderboard and the benchmark description at arcprize.org/arc-agi/3.
Does It Tell You Which Model to Run?
Partly, and less than you would like. ARC-AGI-3 is a good signal for exploration under uncertainty — an agent handed an unfamiliar system and asked to work it out. It is a weak signal for most of what a personal or team agent does day to day, which is well-specified work in a familiar environment: reply to this, summarise that, run the scheduled job, call the API you already configured.
For that everyday work, cost per token, latency, tool-calling reliability and context length matter more than frontier reasoning scores. A practical setup is a cheap, fast model as the everyday driver with a frontier model reserved for genuinely hard steps — which is what runtime model switching is for. Our best models guide works through the trade-off with the model line-up as it stands.
Related Guides
- Best models for OpenClaw — choosing by workload rather than by leaderboard
- Best models for Hermes Agent
- Multi-model switching
- OpenClaw + GPT-6 Astra
- Claude Fable 5.1
- Compare all models