ARC Prize 2026 (ARC-AGI-3) on Kaggle: Deadline, Rules and Our Run
ARC Prize 2026 – ARC-AGI-3 is a Kaggle competition with an $850,000 prize pool. Agents play interactive games they have never seen and are scored on levels completed and action efficiency. Submissions closed on 2 November 2026 and final results are due on 4 December 2026. Each submission is a notebook that runs for up to 9 hours on an RTX 6000 GPU with no internet.
Competition facts
- Host
- ARC Prize Foundation, on Kaggle (Featured)
- Prize pool
- $850,000
- Submissions closed
- 2 Nov 2026, 23:59 UTC
- Paper track due
- 8 Nov 2026
- Final results
- 4 Dec 2026
- Task
- An agent plays interactive game environments it has never seen
- Score
- Levels completed and action efficiency, on hidden games
- Runtime
- Notebook, RTX 6000 (48 GB), no internet, up to 9 hours
- Medal lines (public)
- Bronze: rank 492 or better · silver: 32.97 or better
- Teams
- 4,926 on the public leaderboard (final)
- Leader
- 75.56 public
Deadline, prize and team count from the Kaggle API; rules from the competition page. Check the official rules before you rely on any of it.
Our run with ResearchForge
- 313 / 4,926
- Public rank
- 32.51
- Best public score
- 6.4%
- Top of the field
- 14
- Experiments · 10 rejected
🥉 Medal standing · Bronze-medal range
Rank 313 is inside the bronze-medal range on the public leaderboard (top 10% is rank 492 or better) and 0.46 below the silver line at 32.97. Medals are decided on the hidden games on 4 Dec, so this can still move.
From a first score of 0.08 to 32.51. The record shows what really moved it — and the 10 ideas that did not.
- 1–5 Aug0.08
First agents: probe-based and graph-frontier exploration
- 10 Aug1.21
Agent harness + Qwen3.6-27B served with vLLM
- 25 Aug1.94
LoRA fine-tune on 9 of the agent’s own winning games — validated
- 1 Oct21.06
Rebuilt on a strong public notebook by dfranzen
- 30 Oct32.51
Our best public score, on our own variant of that pipeline
- 2 Nov313th
Submissions closed — rank 313 of 4,926, inside the bronze-medal range
- 4 Dec—
Final results on the hidden games
What the record showed
Two ideas that scored higher on our offline benchmark did worse on the hidden games. Both were rejected on the evidence, not kept on a hunch.
Questions people ask
When did ARC-AGI-3 submissions close?
Submissions for ARC Prize 2026 – ARC-AGI-3 on Kaggle closed on 2 November 2026 at 23:59 UTC. Paper track entries were due on 8 November, and final results on the hidden games are due on 4 December 2026.
What is the ARC-AGI-3 runtime limit?
Each submission is a Kaggle notebook that runs for up to 9 hours on an RTX 6000 (48 GB) GPU with no internet access, so the agent must use a model that runs locally inside that limit.
How is ARC-AGI-3 scored?
Agents play interactive game environments they have never seen. The score reflects levels completed and how efficiently the agent acts. The public leaderboard uses one set of games and the final standings use hidden games, so ranks can move when results are announced.
How did ResearchForge do on ARC-AGI-3?
Our best public score was 32.51, up from a first score of 0.08, ranking 313 of 4,926 teams on the public leaderboard — inside the bronze-medal range (top 10%) and 0.46 below the silver line. Medals are decided on the hidden games on 4 December 2026. ResearchForge recorded 14 experiments; 10 were rejected with a written reason.
Sources
- Kaggle competition page
- ARC Prize 2026 – ARC-AGI-3 (arcprize.org)
- Our scores and ranks: public Kaggle leaderboard, 15 Nov 2026, and the ResearchForge project record.
Run it on this competition
Kaggle Overnight is open-source autoresearch for Kaggle: it keeps your score level with the best public notebooks, then keeps experimenting on top of them. Give it the competition URL tonight.