Why 9x9 Go instead of the full 19x19 board?

The paper says 9x9 gives a strong signal at about a quarter of the cost and time of 19x19, and that the 9x9 board is still far from solved. The author plans 13x13 and 19x19 once models saturate 9x9.

Can a model look up the best moves?

No. In Track 1 the model has no tools and is never allowed to call KataGo. It sees only the board, the recent moves and the list of legal moves.

Why are the intervals so wide?

Each model plays 14 to 30 games, because every move is a paid API call. Fewer games mean wider intervals, so models within about 200 Elo of each other can swap places.

What do GoBench Tracks 2 and 3 measure?

In Track 2 a coding agent spends 1 to 8 hours writing a Go engine, and the engine plays. In Track 3 an agent with KataGo access writes a 5,000-token tutorial that a model reads before playing without tools. We show Track 1 only, because the other tracks have results for one to three models so far.

Why is GoBench not in the LLMBenchmark Index?

With 14 to 30 games per model the order is not firm enough to move an index. We show GoBench beside the reasoning rankings instead.

GoBench leaderboard

GPT-6 Astra leads: 2567
models measured: 18
lead over #2: +188 Elo
median: 1818

Leader on GoBench · read November 5

1 GPT-6 Astra2567

On November 5, 2026, 18 of our catalog's models had a GoBench result, and OpenAI GPT-6 Astra led with 2567, ahead of OpenAI GPT-6.1 Sol at 2379.

GoBench has each model play full games of 9x9 Go turn by turn, with no tools, against a ladder calibrated with the KataGo engine. It tests multi-step reasoning in a game the model cannot look up.

How to read the score: Scores are Elo ratings. The strongest KataGo reference rates 4,424, so the best models are still far from engine level. The intervals are wide (often ±130 to ±220 Elo), so models within a few dozen points of each other are effectively tied.

18 models with a result

Original board: GoBench (Roland Gao)

GoBench results for models in the catalog
#	Model	Result	Cost per move	Interval (±)	Seconds per move	Coding index	Output $/Mtok
1	GPT-6 AstraOpenAI	2567	$0.15	±140.0	1m 06s	85.1	$50.00
2	GPT-6.1 SolOpenAI	2379	$0.029	±202.0	1m 12s	82.3	$10.00
3	Gemini 4 ArgonGoogle	2353	$0.054	±172.0	49 s	81.0	$10.00
4	Claude Opus 5.5Anthropic	2276	$0.021	±134.0	7.1 s	88.9	$20.00
5	Claude Fable 5.1Anthropic	2269	$0.37	±219.0	1m 02s	84.3	$50.00
6	Claude Opus 5Anthropic	2256	$0.30	±132.0	1m 30s	81.6	$25.00
7	Claude Sonnet 5.5Anthropic	2086	$0.017	±174.0	7.6 s	87.3	$10.00
8	GPT-6 SolOpenAI	2006	$0.031	±222.0	33 s	72.3	$10.00
9	GPT-5.6 SolOpenAI	1929	$0.081	±207.0	1m 19s	70.7	$20.00
10	DeepSeek V4.1 FlashDeepSeek	1707	$0.022	±206.0	2m 20s	56.6	$0.96
11	DeepSeek V4.1 Flash 1017DeepSeek	1634	$0.021	±195.0	2m 50s	70.3	$0.60
12	Gemini 3.8 FlashGoogle	1616	$0.025	±206.0	13 s	61.6	$3.75
13	DeepSeek V4 Flash 0731DeepSeek	1449	$0.016	±227.0	2m 40s	51.8	$0.30
14	Muse Spark 1.3 ContributorMeta	1449	$0.002	±227.0	46 s	Not benchmarked	$0.20
15	Gemini 3.6 FlashGoogle	1336	$0.02	±223.0	5.3 s	62.5	$3.75
16	GPT-6 LunaOpenAI	1120	$0.0018	±191.0	36 s	57.4	$0.50
17	GPT-5.6 LunaOpenAI	1091	$0.0024	±190.0	17 s	55.3	$1.20
18	Grok 4.6xAI	1044	$0.05	±189.0	35 s	65.7	$6.00

Only models in our catalog are listed, and each appears once with its best published run. 18 rows were read from the source on November 5, 2026, and rows that match no catalog model are left out. Cost per move is what GoBench reports one move cost through the model's API, in US dollars. Every other column is a figure the board publishes for the same run.

Data: GoBench by Roland Gao.

How a run works

Each turn the model gets one text prompt: the rules, the 9x9 board drawn as a grid, the last five moves, any moves banned by ko, and the full list of legal moves. It must answer with one entry from that list.
If the answer is not a legal move, the model is asked again with the same prompt until it is. The model has no tools: no code, no search and no Go engine.
The opponents are KataGo players of known strength, rated from 320,000 games among themselves. The ladder runs from a random player fixed at 0 Elo to KataGo with tree search at about 4,400 Elo.
Each model plays 14 to 30 games. Every pairing is played twice with colours swapped, White gets 7 points of komi, and the rating model corrects for the colour that moves first.
Results go into a Bradley–Terry Elo model. A 400-point gap means 10-to-1 odds of winning, and an 800-point gap means 100-to-1.

At a glance

Run by: Roland Gao, independent AI researcher
Task: Play full games of 9x9 Go, one move per turn, without tools
Rules: Tromp-Taylor area scoring, positional superko, 7 komi
Shown here: Track 1 (models without tools)
Games per model: 14 to 30
Scale: random player 0 Elo; KataGo policy network with no search about 3,330; strongest KataGo about 4,424
In our indexes: No, shown beside the reasoning rankings

The best model on this board, OpenAI GPT-6 Astra, is 763 Elo below KataGo's policy network playing with no search at all.

What the source reports

GoBench Elo tracks ARC-AGI closely: the paper reports a Pearson correlation of 0.83 with both ARC-AGI-1 and ARC-AGI-2 semi-private scores.
Keeping the conversation across turns pays: for GPT-6 Astra at High, multi-turn context cut time per move by a factor of four and cost by 40% while keeping the same Elo.
When models may write code instead of playing move by move (Track 2), the Go engines they build rate well above any model playing without tools.
The paper names two checks it added after inspecting games: its own superko checker, because the game interface accepted illegal repeats, and a check for KataGo players that pass while dead stones remain on the board.

Role in the index

GoBench is shown as published and does not feed any of our indexes. Read the methodology

Questions

Why 9x9 Go instead of the full 19x19 board?: The paper says 9x9 gives a strong signal at about a quarter of the cost and time of 19x19, and that the 9x9 board is still far from solved. The author plans 13x13 and 19x19 once models saturate 9x9.
Can a model look up the best moves?: No. In Track 1 the model has no tools and is never allowed to call KataGo. It sees only the board, the recent moves and the list of legal moves.
Why are the intervals so wide?: Each model plays 14 to 30 games, because every move is a paid API call. Fewer games mean wider intervals, so models within about 200 Elo of each other can swap places.
What do GoBench Tracks 2 and 3 measure?: In Track 2 a coding agent spends 1 to 8 hours writing a Go engine, and the engine plays. In Track 3 an agent with KataGo access writes a 5,000-token tutorial that a model reads before playing without tools. We show Track 1 only, because the other tracks have results for one to three models so far.
Why is GoBench not in the LLMBenchmark Index?: With 14 to 30 games per model the order is not firm enough to move an index. We show GoBench beside the reasoning rankings instead.