125 billion parameters. One NVIDIA card.

Agents that learn unfamiliar rules, on a fixed machine.

WHAT IT IS

ARC-AGI-3 is the ARC Prize 2026 benchmark for interactive reasoning. An agent is placed in a game it has never seen, with no instructions, and is scored on how few actions it needs compared with people. Every team ran on the same machine: one NVIDIA RTX PRO 6000, nine hours, 110 hidden games, no internet.

WHY IT MATTERS

It measures how quickly an agent works out the rules of something new, not what it has memorised. Because the hardware is fixed, inference efficiency is part of the result: every token the card does not waste is another move an agent can think through.

WHAT WE DID

We built an agent harness in which the model's only tool is a Python sandbox, with notes that survive context trimming and checked play. We built the serving system that cut the model to fit the card and ran thirty-two agents on it at once. Keith Tyser built a trainer that works inside the same single card. The harness builds on Tufa Labs' public harness and Daniel Franzen's public Milestone 2 notebook.

DEMONSTRATED

5th of 4,923teams on the public leaderboard, score 56.02

605 → 2,150tokens per second over a whole run on one card, from the public notebook we started with to our final build

10 → 32agents running at once on that card

3 of 12changes that improved in our local tests also improved on the public leaderboard

WHAT DID NOT HOLD UP Of twelve changes that improved in our local tests, only three also improved on the public leaderboard, and both reinforcement-learning adapters we submitted scored lower than the package they were added to. The evidence notes list all fourteen comparisons.

WHERE TO INSPECT

Models that write the agents: a testbed for recursive self-improvement.

WHAT IT IS

A model reads the current leading agents, writes a new one as code, tests it against the best, and submits it to a leaderboard. The next round starts from what the last one found. Agents compete head to head in games, or on a benchmark where the task has an exact score. Every agent's source code, and the model's own working log, are public.

WHY IT MATTERS

Recursive self-improvement is hard to study because it is hard to measure. The arena makes one loop small enough to observe: each cycle is recorded, every result is scored by the same fixed rules, and the code that produced it can be read. It supports a plain question: do improvements compound, and where do they stop?

WHAT WE BUILT

For each task: a task definition and an instruction file (program.md) that tells the model how to improve, hand-written seed agents as the baseline, a sandbox that runs the code the model submits, and a scoring rule. Games use ratings from head-to-head matches; algorithm and mathematics tasks use exact benchmark scores. A heartbeat log records each cycle, and anyone can play the agents in the browser.

DEMONSTRATED

36,000evolution cycles across 19 tasks, March to October 2026

1.46 millionmatches and benchmark runs scored

4,716agents listed in the archive, up to 500 per task, mostly written by Claude Haiku 4.5

20 tasksboard games, algorithms, mathematics and chip design

WHAT IT HAS NOT SHOWN YET The loop improves agents; it does not yet improve itself. The instruction file that steers each loop (program.md) was versioned but revised on only one task, and on sorting the built-in baseline still tops the ratings. These are the next results to earn.

PUBLIC SOLUTIONS · THE LEADING AGENT ON EACH TASK, WITH ITS SOURCE CODE

Each agent's name opens a readable page with its source code, copied as the arena stores it. The raw links return the same code and the model's own working log as JSON from the arena's open API. The leader is the highest-rated agent a model wrote, or, on benchmark tasks, the first on the task's own board.

WHERE TO INSPECT