ARC Prize 2026 is a $2 million open competition run by the nonprofit ARC Prize Foundation, with final submissions due on November 2, 2026, papers due November 8 and results announced December 4. It has three tracks: ARC-AGI-3, an interactive agent challenge; ARC-AGI-2, a static reasoning challenge; and a Paper Prize. Every prize requires an open-source, reproducible solution submitted through Kaggle, with no internet access during evaluation. With the deadline three days away, here is how the contest works and how it compares with other big AI prizes.

What ARC measures

ARC-AGI is a family of benchmarks designed to be easy for people and hard for AI. The foundation, co-founded by Mike Knoop and François Chollet, defines AGI as “a system that can match the learning efficiency of humans,” and argues that scaling alone will not get there. The tasks avoid reliance on memorized knowledge: a solver must infer a rule from a few examples or, in the newest version, discover how an environment works by playing it.

ARC-AGI-2, introduced in 2025, builds on the original ARC-AGI-1 format used from 2020 to 2024. Each task shows grids of colored cells; the solver must produce the output grid for a new input. For scoring, you submit exactly two predictions per test input, and a task counts as solved if either matches the ground truth. The final score is the average across all test outputs.

ARC-AGI-3 is different in kind. The foundation calls it the first interactive reasoning benchmark: agents are dropped into novel environments with no instructions and must explore, learn goals on the fly, build a world model and plan. All environments are described as solvable by humans, and a 100% score means an agent beats every game as efficiently as a person. The official page lists four capabilities under test: exploration, modeling, goal-setting, and planning and execution.

The three tracks and their prizes

ARC-AGI-3 — $850,000 in total. A $700,000 Grand Prize goes to the first eligible agent that scores 100% on the evaluation; if nobody wins, it rolls into the next annual competition. A guaranteed $75,000 Top Score Award pays $40,000, $15,000, $10,000, $5,000 and $5,000 to the top five. Another guaranteed $75,000 went to milestone prizes for open-sourcing by June 30 and September 30, 2026 ($25,000, $10,000 and $2,500 for the top three each time).

ARC-AGI-2 — $700,000 in total. The goal is 85% accuracy on the private evaluation set within Kaggle’s efficiency limits. The $275,000 Progress Prizes pay eight places, from $75,000 for first to $15,000 for seventh and eighth. A $275,000 Grand Prize goes to the best Solution Writeup, scored equally on six criteria from 0 to 5, and all artifacts must be open sourced and attached within seven days of the submission deadline. A $150,000 Bonus Prize goes to the first solution that scores at least 85%; if unclaimed, the foundation intends to roll it into 2027.

Paper Prize. Awards for papers that advance understanding of how to reach strong performance on ARC-AGI. The overview page does not list amounts in the text we could read, so check the Kaggle page before planning around it.

Rules that shape what you can build

  • Open source or nothing. Code and methods you author must be released under a permissive public-domain-style license such as CC0 or MIT-0. Third-party components need at least an open license that permits sharing, such as Apache-2.0 or GPLv3. Leading participants must publish before getting official private scores.
  • Kaggle notebooks only. Submissions go through the designated Kaggle competition for each track.
  • Offline evaluation. There is no internet access during scoring, so API-based systems such as GPT or Claude cannot be called. You must bring your own model weights or algorithm.
  • Compute limits. Hardware and runtime limits are set per competition on Kaggle.
The Kaggle logo
ARC Prize 2026 submissions are made through Kaggle notebooks; the logo identifies the platform. Image: Kaggle / Wikimedia Commons, Public domain

These constraints are deliberate. The foundation’s stated mission is open progress toward AGI, which is why prize money tracks transparency, not just the leaderboard score. For general advice on tactics and pitfalls, our guide to winning at Kaggle and AI hackathons applies, and our explainer on how AI competitions are judged explains why the private test set matters.

How it compares with AIMO

The AI Mathematical Olympiad (AIMO) Prize, funded by XTX Markets, is a $10 million challenge fund for open AI models that reason mathematically. Its $5 million grand prize goes to the first publicly shared model that, in an approved competition, performs at the standard of an International Mathematical Olympiad gold medal, with up to $5 million more in progress prizes. Progress prizes opened in April 2024, October 2024 and November 2025, according to the AIMO site. Both contests share an open-source ethos, but AIMO targets one skill, olympiad mathematics, while ARC deliberately targets learning efficiency on novel problems. See also our history of famous AI competitions.

Practical tips if you still want to enter

  1. Register on Kaggle and read the exact rules for the track you want; the two competition pages above link to them.
  2. Play the public ARC-AGI-3 games yourself first. The foundation offers a public game set, human leaderboards and an SDK.
  3. Start from the foundation’s technical guide for ARC-AGI-1 and ARC-AGI-2, which explains data structure and common approaches.
  4. Budget time to open-source your code before the deadline; ineligible entries are removed from prize contention.
  5. Even if you miss November 2, you can use public data and leaderboards independently.

What to watch next

  • The November 2 submission deadline and November 8 paper deadline.
  • Results on December 4, including whether anyone reaches the 85% ARC-AGI-2 bonus threshold.
  • Whether any agent comes close to 100% on ARC-AGI-3, or the Grand Prize rolls over.
  • The next AIMO progress prize announcements.

Key takeaways

  • ARC Prize 2026 offers $2 million over ARC-AGI-3, ARC-AGI-2 and a Paper Prize.
  • Final submissions are due November 2, papers November 8, results December 4.
  • All prizes need open-source, reproducible Kaggle submissions that run offline.
  • ARC-AGI-3 tests interactive exploration; ARC-AGI-2 tests static grid puzzles with two attempts per input.
  • AIMO is a separate $10 million fund aimed at olympiad-level math.

Frequently asked questions

What is ARC Prize 2026?

It is a $2 million competition from the ARC Prize Foundation to advance open AI research, with three tracks: ARC-AGI-3, ARC-AGI-2 and a Paper Prize.

When is the ARC Prize 2026 deadline?

Submissions are due November 2, 2026, papers on November 8, and results will be announced on December 4, 2026.

Can I use ChatGPT or Claude in my entry?

No. Evaluation runs on Kaggle without internet access, so API-based systems cannot be called, and your solution must be open source.