ARC-AGI-3 run log

Played through 2026-10-11hi@nodecode.ai

Nodecode on GPT-6.1 Sol at xhigh won all 25 public games, a mean score of 98.89, for $38.50 at list price over 2,694 model calls. This page holds every game, every model call and every action behind that line on the chart. The games are deterministic, so the actions replay to the same scores in ARC Prize’s own toolkit, and the script that does it is here too.

The 25 games

cost over every attempt

Game	Result	Levels	Actions	Score	Calls	Cost
ar25	WIN	8/8	264	100.00	47	$0.40
bp35	WIN	9/9	698	86.35	207	$2.87
cd82	WIN	6/6	97	100.00	32	$0.29
cn04	WIN	6/6	219	100.00	88	$1.33
dc22	WIN	6/6	544	100.00	170	$2.54
ft09	WIN	6/6	87	100.00	34	$0.27
g50t	WIN	7/7	387	95.91	90	$1.62
ka59	WIN	7/7	703	90.01	100	$1.06
lf52	WIN	10/10	843	100.00	250	$4.17
lp85	WIN	8/8	97	100.00	53	$0.91
ls20	WIN	7/7	459	100.00	134	$1.89
m0r0	WIN	6/6	250	100.00	64	$0.69
r11l	WIN	6/6	99	100.00	88	$1.19
re86	WIN	8/8	709	100.00	175	$2.55
s5i5	WIN	8/8	355	100.00	87	$1.28
sb26	WIN	8/8	126	100.00	24	$0.21
sc25	WIN	6/6	191	100.00	56	$0.59
sk48	WIN	8/8	563	100.00	431	$7.45
sp80	WIN	6/6	144	100.00	65	$1.24
su15	WIN	9/9	148	100.00	107	$1.02
tn36	WIN	7/7	185	100.00	68	$0.85
tr87	WIN	6/6	162	100.00	61	$0.56
tu93	WIN	9/9	193	100.00	85	$1.18
vc33	WIN	7/7	227	100.00	88	$0.97
wa30	WIN	9/9	638	100.00	90	$1.36

155,316,430 tokens in all. The score is each game’s scorecard score; the mean over the 25 is 98.89.

Attempts not kept

$7.60 of the total

The first sweep ran out of API credit with four games unfinished and one not started. Those five were resumed from their last position, or started again. A resume continues the same session, so its calls are counted once. The attempts below were dropped, and their cost is in the total: without them the kept runs cost $30.90.

Game	Attempt	How it ended	Calls	Cost
lf52	gL25	Stopped when API credit ran out	60	$0.73
re86	gJ25	Stopped when API credit ran out	104	$1.60
sk48	gJ25	Stopped when API credit ran out	125	$2.77
sk48	gL25	Stopped when API credit ran out	50	$0.65
sk48	gM25	Hit the 1,000-action cap at 4 levels	105	$1.85

Check it yourself

verify.py makes each game fresh in ARC Prize’s toolkit, plays the logged actions, and compares the end state, levels, action count, scorecard score and a SHA-256 of the final frame with games.csv. Point it at the toolkit’s folder of the 25 public games. In its default mode the toolkit downloads a game there the first time it is made, with an API key from arcprize.org.

pip install arc-agi==0.9.9
python verify.py --games path/to/environment_files
...
all games replay to the logged result

For the cost, list_usd in calls.csv sums to $38.50. Each row is one model call with its uncached, cached and output tokens, priced at the list rates below.

The run

	Setting
Harness	Nodecode, build c49352e35 (2026-10-08), with its ARC game tools. One autonomous session per game.
Model	GPT-6.1 Sol, reasoning effort xhigh, 180,000-token context window
Games	The 25 ARC-AGI-3 public games in ARC Prize's toolkit, arc-agi 0.9.9, offline, seed 0
Limits	1,000 actions and 4 hours per game. The fresh sk48 rerun had 1,500 actions and won in 563.
Score	The toolkit's scorecard: each level against its human baseline, levels weighted by their index, a won game capped at 100
Prices	OpenAI's list rates for GPT-6.1 Sol: $2.50 per million input tokens, $0.10 cached, $10 output
Played	2026-10-09 to 2026-10-11

Model transcripts are not published.

Files

File	What
nodecode-arc-agi-3-gpt-6.1-sol-xhigh.tar.gz	All of the below in one archive
games.csv	One row per game: the attempt kept, end state, levels, actions, score, per-level actions against the human baseline, final frame hash, calls and dollars over every attempt
attempts.csv	One row per attempt, with the calls and dollars it added and whether it is the kept run
calls.csv	One row per model call: tokens by kind and the price at list
actions/ar25.jsonl	The kept run's actions for a game, in order (one file per game, named by game)
verify.py	Replays every game in ARC Prize's toolkit and compares it with games.csv
README.md	How the run was set up and how to check it