TN36 / Level 6
85.0% fewer actions
32 OY1 actions / 214 Provider Adapter actions
Research / ARC-AGI-3
25 games.
183 levels.
All complete.
A 100.0 score.
With the work to back it up.
The OY1 research harness completed every level across all 25 public ARC-AGI-3 games. Explore the run, compare token usage, and replay the decisions.
- Public-set score
- 100.0
- Games completed
- 25 / 25
- Levels completed
- 183 / 183
- Successful-run inference cost
- $415.37
Completed 9 September 2026 · GPT-6 Astra, high reasoning · 6,732 environment actions. Published run accounting
Same public games / Same model setting
The same finish.
83.03% fewer tokens.
Both published runs completed all 183 levels. The OY1 harness reported 153.06 million total tokens, compared with 901.97 million in the Provider Adapter public replays.
Provider Adapter901.97M
OY1 harness153.06M
25 exact public game versions. GPT-6 Astra with high reasoning in both runs. Zero-based scale; totals include compaction where reported.
View the underlying dataThis is a comparison of published runs, rather than a controlled component ablation or a claim about dollar-cost savings. The public games were used during development; held-out performance has not been measured. The scorecard is hosted by ARC Prize; the result has not been independently verified by ARC Prize. The $415.37 figure covers the completed successful run, excluding development experiments.
Replay the work
The decisions.
Not just the score.
Inspect full game recordings on ARC Prize. These two level highlights show how the published runs reached the same outcome with different action counts.
LF52 / Level 10
54.5% fewer actions
50 OY1 actions / 110 Provider Adapter actions
Open source / Open methods
Results you can read.
Methods you can inspect.
The benchmark is one concrete experiment in our broader research into reasoning, memory, agent interaction, and AI orchestration.
01 / Results
The complete run.
Per-game results, action counts, runtime, token accounting, and successful-run inference cost.
02 / MethodInside the harness.
How observation, prediction checks, evidence, and memory guide the next action.
03 / ReproduceOpen to inspection.
Source, configuration, and instructions for reproducing the published experiment.
Looking for the original evaluation page? Read the evaluation and scope
A separate integration example
See the loop
in practice.
A short MCP demonstration walks through prediction, checking, evidence, and memory on a recorded task.
These public artifacts document a specific harness and experiment. OY1 is our full commercial AI orchestration platform and API, spanning the breadth of AI work: research, software, automation, images, video, and connected actions.
Your next chapter starts here
A question worth
exploring?
Talk with us about research and collaboration.