AGI guide · Beginner · 10 min
ARC Prize: the competition, its rules, editions, results, and lessons
ARC Prize rewards open methods for solving ARC-AGI tasks. Learn how its rules work, what past editions achieved, and what those scores can actually tell us.
ARC Prize is a research competition that rewards open, reproducible solutions to ARC-AGI tests, along with research explaining the methods behind them. Its value extends beyond the leaderboard: it helps researchers study how systems discover rules, learn during a task, and use their resources. As of October 24, 2026, the 2024 and 2025 editions have concluded; the 2026 edition is still running.
ARC Prize and ARC-AGI: the competition and its measuring tool
ARC-AGI is the test; ARC Prize provides a framework for organizing research around it. Keep that distinction in mind when reading an announcement: solving public tasks, receiving a controlled evaluation, and winning an award are separate events.
In static tasks, a system receives examples of input and output grids. It must infer the transformation, then produce the output for a new input. The challenge lies in discovering the rule, rather than following detailed instructions. Our ARC-AGI reference page explains the test and its different generations.
François Chollet sets out the foundations of this approach in On the Measure of Intelligence, published in November 2019. He distinguishes acquired skill from the efficiency with which a system acquires a new skill. From this perspective, measuring intelligence requires accounting for prior knowledge, available experience, and the difficulty of generalization.
The competition turns this scientific question into an experimental problem. A team proposes a method; an evaluation measures its behavior; publication lets other teams examine what works. Generalization remains the research objective, while the score provides an operational measure.
This distinction prevents a common misunderstanding: ARC Prize does not issue a universal certification of intelligence. It organizes comparisons under specific conditions. To place this research alongside other approaches, see our AGI and research overview.
What a submission actually has to do
Consider an illustrative example that does not reproduce any official task. Several grids show colored objects and the results of a transformation. One possible rule might move each object to an edge; another might complete a symmetrical pattern. A solution must distinguish between these possibilities using the available examples.
A convincing method must explain every example, then work on the input whose answer is hidden. Finding a transformation that fits a single example is not enough: several rules may produce the same apparent result.
When examining a submission, distinguish between the operations involved:
- representing grids, objects, or relationships;
- proposing possible transformation rules;
- checking whether they fit the examples;
- selecting an answer to submit;
- meeting the compute and formatting requirements.
This breakdown helps you analyze a system without making assumptions about its architecture. One method may generate a grid directly; another may build a program that generates it. The visible result alone does not reveal the process used to reach it.
The concept of a benchmark is useful here: a test defines a procedure for observing performance. You need to examine that procedure before interpreting the result. Our article ARC-AGI explained offers a shorter introduction to how the puzzles work.
As you read, focus on what information the system had before answering and what operations it was allowed to perform. These questions remain relevant even when scores improve substantially.
Completed editions: results and valid comparisons
The results of completed editions provide historical reference points. They should always be presented alongside the test used: percentages from different generations do not represent the same level of difficulty.
| Criterion | 2024 edition | 2025 edition |
|---|---|---|
| Competition test | ARC-AGI-1 | ARC-AGI-2 |
| Top team on the score leaderboard | the ARChitects | NVARC |
| Published private score, rounded | 53.5% | 24.0% |
| Edition’s grand prize | Not awarded | Not awarded |
| Other form of recognition | Research paper awards | Research paper awards |
The 2024 figures come from the official ARC Prize 2024 results, accessed on October 24, 2026. For 2025, the official analysis published on December 5, 2025 reports private ARC-AGI-2 evaluation scores of 24.03% for NVARC, 16.53% for the ARChitects, and 12.64% for MindsAI.
The apparent drop in the winning score between these editions does not demonstrate a decline in capability. The test changed: comparing percentages without identifying the underlying test is like comparing grades from different exams.
You also need to distinguish the best submission in a particular edition from the best result later achieved on the same benchmark. Competition results remain historical records tied to a closing date and a set of rules.
Finally, topping the score leaderboard does not automatically mean reaching the main objective. Intermediate awards exist precisely to recognize progress while the target threshold remains out of reach. This distinction gives the competition continuity: an edition can produce useful methods without solving the entire problem.
What the award-winning research contributed
The research awards offer another way to understand the competition. They draw attention to an explanatory idea, its reproducibility, and its potential for reuse, beyond the question of who finishes first.
The 2024 awards recognize, among others, work by Li and coauthors combining induction and transduction, and work by Akyürek and coauthors on test-time training. In 2025, the recognized approaches include Alexia Jolicoeur-Martineau’s recursive networks; evolutionary program synthesis by Julien Pourcel, Cédric Colas, and Pierre-Yves Oudeyer; and work without pretraining by Isaac Liao and Albert Gu.
The official presentation of the 2025 winners describes these research directions and the leading competition solutions. Their diversity cautions against treating all progress as the same process: updating parameters, refining an answer, and searching for a program are different mechanisms.
To understand them, ask concrete questions. Does the system revise its representation of the problem? Does it try several solutions? Does it learn a transformation specific to the task? Does it retain knowledge that can be reused?
The methodological lesson is that the final answer alone cannot explain success. You need to examine what changes during problem-solving and what was already present before it began.
This does not automatically identify a winning architecture for general intelligence. A technique that works well on grids still needs to demonstrate its value elsewhere. The competition provides experiments and research directions; their broader significance depends on what other researchers can verify.
In 2026, several tracks have distinct objectives
The 2026 edition broadens the competition’s scope. The official ARC Prize 2026 schedule, accessed on October 24, 2026, announces a $2 million prize pool and three tracks: ARC-AGI-2, ARC-AGI-3, and research papers.
The launch date is March 25, 2026. Code submissions are due on November 2, papers on November 8, and results on December 4, 2026. These deadlines explain why this page does not name any final winners for the edition.
The 2026 ARC-AGI-2 track targets an 85% success rate on its private evaluation within the competition’s efficiency limits. It announces $700,000 divided among progress awards, assessment of a solution paper, and a bonus conditional on reaching the threshold.
That allocation deserves attention: the term “grand prize” does not necessarily carry the same conditions across tracks or years. For ARC-AGI-2 in 2026, the page distinguishes the solution paper award from the bonus tied to the target score.
When reviewing a prize amount, check which category it funds, what triggers payment, and whether it is guaranteed. An announced prize pool does not necessarily mean that the full amount will be distributed.
If you plan to enter, consult the track rules and the submission platform’s schedule. An overview explains how the competition is organized; it does not replace the detailed requirements that apply to your entry.
The rules: openness, reproducibility, and resources
The overall 2026 framework requires open, reproducible solutions to qualify for awards. It also specifies a Kaggle evaluation without internet access, so a submission cannot call a remote API during that run.
Openness should make a method available for examination. To analyze it, a reader needs to identify the components used, the data transformations performed, and the operations required to reproduce the result. A percentage accompanied by a vague description leaves too many questions unanswered.
Resources are another dimension of a result. Consider these hypothetical setups:
| Criterion | Locally evaluated system | System using a remote service |
|---|---|---|
| Execution | Components available in the environment | Depends on an external service |
| Resources to document | Hardware, memory, runtime | Calls, remote compute, runtime |
| Reproduction | Depends on the artifacts provided | Also depends on access to the service |
| Competition eligibility | Must be checked against the rules | Incompatible with an offline evaluation if remote calls are required |
This table does not rank the approaches by quality. It shows why a result obtained in one setup cannot automatically be carried over to another.
Before comparing teams, check the time allowance, hardware, additional data, and publication requirements. Do not assume that limits from an earlier edition still apply.
A useful comparison keeps these constraints alongside the score. You can then distinguish an improvement in method from an increase in available resources, or make clear that both have changed.
ARC-AGI-3 adds learning through interaction
The 2026 ARC-AGI-3 track asks agents to interact with unfamiliar environments without being given instructions. It emphasizes exploration, building a model of the environment, identifying goals, and planning.
Its announced prize pool as of October 24, 2026, is $850,000. This includes a grand prize conditional on a 100% score, leaderboard awards, and milestone prizes. The threshold belongs to the evaluation protocol; it should not be presented as a declaration of AGI.
Interaction changes the kinds of errors worth examining. In a static puzzle, you mainly observe an answer. In an environment, you can also examine whether the system learned anything from its actions.
An illustrative example makes the difference clearer. An agent encounters an element whose effect it does not know. It can act to test a hypothesis, observe the change, and then adjust its strategy. An action that does not immediately achieve the goal may still be useful if it reduces uncertainty.
Conversely, repeating an action without using its outcome may reveal a failure to learn. You therefore need to examine sequences of actions, retained information, and successive decisions.
This perspective brings evaluation closer to a practical question: how does a system become competent in an unfamiliar situation? It does not remove the need to control resources. Spending a long time exploring and learning efficiently remain separate properties.
Model results outside the competition change the picture
A published evaluation can report a high score without being an eligible prize submission. This distinction is especially necessary in 2026, when you encounter scores close to the competition’s thresholds.
On September 3, 2026, ARC Prize reports that GPT-6 Astra scored 62.7% on semi-private ARC-AGI-3 using its standard setup, compared with 99.9% using a provider-adapted setup. The reported total costs for these evaluations are approximately $26,000 and $19,000, respectively. The official evaluation report specifies the conditions and states that the result does not constitute proof of AGI.
The gap shows why a model name alone is not enough to describe an experiment. Retaining information between steps and changing the system surrounding the model calls can substantially alter the behavior observed.
When interpreting an announcement, distinguish between:
- the model used;
- the complete system that enables it to act;
- the task set evaluated;
- the resources consumed;
- the result’s status within the competition.
These distinctions let you recognize progress without attributing an award that has not been announced. They also prevent you from treating a semi-private score as though it came from the competition’s final private evaluation.
Our article on how to read model leaderboards develops this approach. The basic habit is straightforward: look for the measurement conditions before comparing the displayed numbers.
What the results allow us to discuss
The scientific question goes beyond choosing between deep learning and symbolic programs. The recognized research explores several ways to represent a problem, search for a solution, and correct it.
Researchers also differ on what success would explain. François Chollet emphasizes efficiency in acquiring new skills. Other contributions recognized by the competition make the case for continuing to pursue deep learning approaches: work by Jack Cole and Mohamed Osman, for example, appears among the 2025 honorable mentions. Strong performance alone does not settle every theoretical question.
To assess an explanation, ask what the experiment isolated. If a team changes the data, architecture, and search budget simultaneously, the overall improvement does not allow the entire gain to be attributed to a single change.
Also examine what transfers. A procedure designed to recognize objects in a grid may contribute a useful idea without demonstrating general capability across documents, social interactions, or physical situations.
These qualifications do not make scores useless. They clarify their purpose: to distinguish systems within a controlled setting, then support more targeted experiments.
You can therefore read each edition as a collection of results and hypotheses. The leaderboard establishes performance; papers propose explanations; reproductions and further evaluations determine how far those explanations hold.
How to keep track of ARC Prize over time
To follow the competition without confusing its different announcements, keep a short record for each result: edition, track, team, score, evaluation set, resources, and publication status. Add the date you checked it when the leaderboard is still changing.
Then distinguish documents by their purpose. The rules determine eligibility; the official results record awards; the research paper explains a method; a model evaluation describes a particular experiment. None of these documents replaces the others.
When a result seems decisive, identify what it lets you compare directly. Two methods evaluated on the same tasks under the same constraints offer a more interpretable comparison than two announcements using different protocols.
For a practical decision, keep your own use case in view as well. Success on ARC-AGI does not directly measure how useful an assistant will be in your work. The Benchmark Lab helps place measurements within their intended scope.
The lasting lessons of ARC Prize come from this disciplined approach to reading results:
- distinguish the test, the competition, and the awards;
- connect every score to its evaluation conditions;
- examine methods and their reproducibility;
- check what transfers beyond the benchmark;
- treat results from an ongoing edition as provisional.
Related pages
Frequently asked questions
What is the difference between ARC Prize and ARC-AGI?
ARC-AGI is a series of tests of reasoning and adaptation. ARC Prize organizes competitions and rewards contributions that use those tests. A published ARC-AGI result is therefore not necessarily an eligible competition entry.
Do you have to reach the main threshold to win a prize?
No. Depending on the edition and track, awards recognize leading scores, intermediate milestones, or research papers. Check the categories for the year concerned, because their requirements change.
Can you enter using a proprietary model API?
The 2026 Kaggle evaluation runs without internet access, so submissions cannot call a remote API. Model evaluations conducted outside the competition follow a different protocol. The rules of your chosen track determine which solutions are eligible.
Does a nearly perfect ARC-AGI score prove AGI?
It demonstrates performance under a particular protocol, using a specific system and set of resources. It does not establish general capability across all situations. You also need to consider task novelty, efficiency, and transfer to other environments.
Are the final ARC Prize 2026 results available?
As of October 24, 2026, the competition is still running. The official schedule sets the results announcement for December 4, 2026. Interim leaderboards and evaluations conducted outside the competition are not its final results.