99.61 on ARC-AGI-3

All 25 public games won. All 183 levels cleared. Claude Code with Opus 5, run on 14 August 2026.

The starting point was Jeremy Berman's arc-code, which scored 96.2 on the same 25 games. This run reuses his design and adds a small number of changes aimed at the specific ways his sessions lost. Those changes are described below, along with everything that went wrong along the way.

What the number means

ARC computes the score on its own servers. You do not calculate it yourself. Each level is scored against a published baseline action count, so clearing a level in fewer actions than the baseline earns full marks and grinding through it earns less. Winning every game is not enough on its own to reach 100.

Of the 25 games, 23 scored a flat 100. Two did not. g50t scored 90.40 because one of its levels took 226 actions against a baseline of 96. bp35 scored 99.80. Everything else was perfect.

Compared with arc-code

Where this sits on the ARC leaderboard

ARC Prize runs a community leaderboard for agent harnesses played on the public games. A score of 99.61 would place third on it, behind Tycho at 100.0 and Retrodict at 99.9, and ahead of baseline1 at 99.0. It would also sit above the Human Intelligence Harness at 95.3, which is people playing the games by hand.

On cost it compares well. Tycho spent $2,986 and Retrodict spent $654. This run spent $572, which is the lowest of the top three.

Was every game won on the first try

Yes, in the sense that matters. No game was ever lost and then played again. The agent never failed a game and got a second chance at it.

What did happen is that ARC dropped the game session for 8 of the 25 games while they were being played. When that happens the game is gone and has to be started again from level one. In those cases the harness restarted the game and let the agent read back its own notes from the interrupted run. arc-code does the same thing when ARC deletes a session.

The important detail is what those interrupted runs had already achieved when they were cut off:

lf52 is the clearest case. The run that got cut off had already cleared all ten levels. The restart was not a second attempt at a puzzle it could not solve. It was re-earning a scorecard for a game it had already beaten, and it took 984 actions the second time against 976 the first, so knowing the answer saved it almost nothing.

What went wrong during the run

Two problems, both outside the games themselves. They are worth writing down because the fixes are part of why the final number came out where it did.

Running out of Claude usage part way through

The Claude account hit its usage limit twice. Sessions started exiting straight away without playing anything and without spending anything. The first time this happened, the harness misread those instant exits as the agent giving up, wound down, and let the shared scorecard go idle. ARC then deleted it, which took the remaining games with it. That first run ended at 63.43 with 12 games never played.

The fix was to notice a session that exits without spending anything, treat it as an infrastructure failure rather than a result, pause every game on one shared timer, and probe again every 20 minutes until the limit clears. Nothing is counted against the game while it waits.

ARC deleting scorecards during those pauses

ARC removes a scorecard after about 15 minutes of no activity, and takes its game sessions with it. While everything was paused waiting for usage to reset, nothing was touching those scorecards, so they were deleted. Worse, the recovery code then restarted those games from level one while the account still could not play, so it spent the progress and got nothing back. Five games were reset this way for no benefit.

Two fixes went in. A background task now touches every unfinished game's scorecard every seven minutes. And a game is never restarted while the account is paused, only when a session can actually play. After that change, a later usage pause happened and no game was reset.

Scorecards still got deleted a few times even so, because polling a scorecard does not appear to count as activity on ARC's side. What actually saved the run was the restart keeping the agent's notes, not the keepalive.

What is in the harness

The design is arc-code's. The agent gets a 15 line description of the environment, a 110 line description of how to play, and one command to act with. It gets no ARC specific knowledge, no solver, no planner, no world model, and no grid tools. It has Bash, Read, Write, Edit, Grep and Glob, and nothing else.

Everything else is written by the model while it plays. Across this run the agents wrote and kept 222 programs totalling about 13,900 lines, which is roughly ten times the size of the harness they can see. Parsers, simulators, search routines and planners, all written during the game and thrown away when the game ended.

The changes that pushed the score up

arc-code published an unusually honest breakdown of how its runs lost, across 191 sessions. Six of the 14 losses were sessions that decided a level was impossible and stopped with between 1,391 and 2,218 of their 2,500 actions unused. In five of those six, the experiment that would have settled the question was already written in the agent's own notes and never run. On one game, no session across 9,792 moves ever pressed one of the seven available buttons. Every change below is aimed at one of those findings.

Whether the changes or the model configuration deserve the credit is not cleanly separable. This run used Opus 5 with the 1 million token context window, while arc-code used its default setting, so some of the difference may come from that rather than from anything in the harness.

Every game

Actions are what this run spent. Baseline is what ARC publishes for the game, added up across its levels. Every scorecard below is public and can be checked.

The games that beat arc-code

Six of the 25 games account for all of arc-code's losses. All six were won here.

• lp85 was the only game they lost outright on a first attempt. Won here in 97 actions against a 388 action baseline.
• lf52 caused half their losses. One of their sessions listed every one of 4,620 states its rules allowed, checked its predictions against 109 recorded moves without a single mismatch, searched 655,440 states, and concluded the level could not be won. Won here, all 10 levels.
• bp35 had a session stop on level 8 after 52 actions on that level, with 2,162 actions left. Won here, all 9 levels.
• sp80 had a session try about 70 million block arrangements in its own simulator, spend 195 actions on the level, call it impossible and stop with 2,218 left. Won here in 158 actions.
• sk48 and wa30 both beat them twice. One wa30 run found the right approach and ran out of moves one step from finishing. Won here, and wa30 took 869 actions against a 1,843 baseline with every level scoring full marks.

Things to be careful about

• These are the 25 public games. They are the development set. Official ARC standings are decided on private tasks under controlled conditions, and nothing here counts toward that.
• The headline model numbers you see for ARC-AGI-3, such as Opus 5 at 30.2, are model only evaluations with no shell and no files. Comparing 99.61 to those is not a fair comparison. It is the same model with a computer in front of it.
• This ran on a normal machine. arc-code ran each game in a locked down container that could only reach three hosts, so lookups were impossible by construction. Here the evidence is behavioural instead. Across 92 session transcripts and 4,351 tool calls there were no web requests, no permission denials, no URLs in any command, and no game ever read another game's files. Good evidence, but not the same guarantee.
• The discovery that a shell and a filesystem take this model from 30.2 to the mid 90s belongs to arc-code. What is new here is the step from 96.2 to 99.61.