Is a GameBench 2 score of 95 the share of games won?

No. It is the share of points a model's programs would be expected to take from DuelLab's fixed reference field of programs. A model at 95 wins nearly every match against that field. DuelLab refits the ratings at every release, so a score can move without the model changing.

Why does one model have several GameBench 2 scores?

DuelLab runs each model at several reasoning settings, such as Low, Medium and XHigh, and ranks each one separately. We show a model's best setting as its result and list the other settings as variants on its page.

What do first-try builds and failed programs mean?

First-try builds is the share of a model's programs that built and passed DuelLab's checks without a repair round. Failed programs is the share that never worked and scored 0. A repair that works is not penalised, so a model can score well with a lower first-try rate.

How is GameBench 2 different from chess leaderboards where models play?

On boards like Kaggle Game Arena the model chooses every move itself. In GameBench 2 the model never plays: it writes a program once, and the program plays hundreds of matches. It measures whether a model can turn the rules of a game into a working, strong engine.

Why is GameBench 2 not in the LLMBenchmark Index?

Its score is relative to DuelLab's reference field and is refitted at every release. Adding it would move our index each time DuelLab publishes, so we show it beside the coding rankings instead.

GameBench 2 leaderboard

Claude Opus 5.5 leads: 94.9
models measured: 86
lead over #2: +0.7
median: 55.7

Leader on GameBench 2 · read November 5

gamebench_2.py


        # GameBench 2, read November 5
        leaders = {
            "Claude Opus 5.5": 94.9,    "Claude Fable 5.1": 94.1,    "GPT-6 Astra": 94.0,
        }

On November 5, 2026, 87 of our catalog's models had a GameBench 2 result, and Anthropic Claude Opus 5.5 led with 94.9, ahead of Anthropic Claude Fable 5.1 at 94.1.

GameBench 2 asks each model to write a program that plays eight strategy games (Go, Chess, Shogi, Xiangqi, Hex, International Draughts, Game of the Amazons and Tumbleweed). DuelLab then plays the programs against each other in deterministic matches.

How to read the score: The score runs from 0 to 100 and shows expected points against a fixed reference field of player programs, averaged across the eight games. It is a relative score, not a percentage of games won; a program that fails to build counts as 0. Small gaps between models are not a firm ordering.

87 models with a result

Original board: DuelLab GameBench 2

GameBench 2 results for models in the catalog
#	Model	Result	Cost per program	First-try builds	Failed programs	Matches	Wins	Draws	Losses	Coding index	Output $/Mtok
1	Claude Opus 5.5Anthropic	94.9	$1.9869	75.0%	0.0%	4,184	3,939	80	165	88.9	$20.00
2	Claude Fable 5.1Anthropic	94.1	$5.1543	75.0%	0.0%	4,180	3,902	82	196	84.3	$50.00
3	GPT-6 AstraOpenAI	94.0	$1.3034	100.0%	0.0%	4,184	3,890	108	186	85.1	$50.00
4	GPT-6.1 SolOpenAI	91.9	$0.2665	100.0%	0.0%	4,184	3,806	99	279	82.3	$10.00
5	Claude Sonnet 5.5Anthropic	91.3	$0.5482	87.5%	0.0%	4,182	3,667	313	202	87.3	$10.00
6	U Pareto 26.10Unbiased	89.7	$0.0296	100.0%	0.0%	4,190	3,625	276	289	Not benchmarked	$3.20
7	Gemini 4 ArgonGoogle	89.0	$0.4405	62.5%	0.0%	4,180	3,628	181	371	81.0	$10.00
8	GPT-6.1 TerraOpenAI	85.7	$0.1068	100.0%	0.0%	4,180	3,408	343	429	77.4	$5.00
9	GPT-6.1 Terra ProOpenAI	84.8	$0.3013	100.0%	0.0%	4,174	3,445	193	536	Not benchmarked	$5.00
10	Claude Opus 5Anthropic	84.7	$1.3797	100.0%	0.0%	6,800	5,497	401	902	81.6	$25.00
11	Kimi K3.1 ProMoonshot AI	84.0	$0.4584	100.0%	0.0%	4,194	3,331	356	507	82.2	$10.00
12	GPT-5.6 SolOpenAI	79.6	$0.7966	87.5%	0.0%	4,178	3,154	309	715	70.7	$20.00
13	GPT-6 SolOpenAI	79.3	$0.2068	100.0%	0.0%	4,184	3,191	254	739	72.3	$10.00
14	Muse WonderMeta	77.1	$0.4876	100.0%	0.0%	4,190	2,934	545	711	72.0	$20.00
15	Claude Fable 5Anthropic	75.0	$1.4706	100.0%	0.0%	4,182	2,944	311	927	75.9	$50.00
16	GLM 5.3Z.ai	72.1	$0.5497	87.5%	12.5%	3,650	2,796	401	453	68.5	$4.40
17	Grok 4.7xAI	72.0	$0.3987	87.5%	12.5%	3,666	2,891	222	553	66.7	$6.00
18	MiMo-V2.6-ProXiaomi	71.3	$0.0569	100.0%	0.0%	4,176	2,764	394	1,018	65.4	$0.87
19	DeepSeek V4.1 Flash 1017DeepSeek	69.7	$0.0755	75.0%	0.0%	4,186	2,594	585	1,007	70.3	$0.60
20	Muse Spark 1.3Meta	68.7	$0.2487	87.5%	0.0%	4,178	2,562	544	1,072	57.4	$4.25
21	Grok 4.6xAI	68.6	$0.1273	75.0%	0.0%	4,188	2,444	803	941	65.7	$6.00
22	Qwen3.8 Max (0902)Qwen	68.1	$0.3986	87.5%	0.0%	4,184	2,401	824	959	67.1	$6.00
23	GLM 5.3 FlashXZ.ai	67.6	$0.0604	100.0%	0.0%	4,184	2,527	519	1,138	Not benchmarked	$1.25
24	DeepSeek V4.1 FlashDeepSeek	67.4	$0.0264	100.0%	0.0%	4,188	2,448	678	1,062	56.6	$0.96
25	Qwen3.8 2.4T A95BQwen	66.7	$0.4318	75.0%	0.0%	4,182	2,349	837	996	Not benchmarked	$6.00
26	Gemini 3.1 Pro PreviewGoogle	64.7	$0.324	62.5%	0.0%	4,182	2,412	551	1,219	32.3	$12.00
27	Gemini 3.8 FlashGoogle	64.2	$0.1432	87.5%	0.0%	4,188	2,435	426	1,327	61.6	$3.75
28	E GlimpseElorian	64.1	$0.1412	75.0%	0.0%	4,182	2,191	946	1,045	Not benchmarked	$3.00
29	Ember-1Fireworks	62.9	$0.2989	87.5%	0.0%	4,186	2,250	725	1,211	52.6	$15.00
30	Gemini 3.6 FlashGoogle	62.8	$0.4249	75.0%	0.0%	8,362	4,425	1,466	2,471	62.5	$3.75
31	Gemini 3.5 FlashGoogle	62.6	$0.2706	62.5%	0.0%	4,188	2,256	646	1,286	39.8	$9.00
32	Claude Opus 4.8Anthropic	62.4	$0.9312	75.0%	12.5%	3,672	2,355	484	833	62.8	$25.00
33	GPT-5.4OpenAI	61.8	$1.519	87.5%	12.5%	3,694	2,342	489	863	64.4	$15.00
34	Qwen3.8 FlashQwen	61.7	$0.0377	62.5%	0.0%	4,188	2,061	961	1,166	Not benchmarked	$0.47
35	Grok 4.5xAI	61.4	$0.0746	75.0%	0.0%	4,178	2,193	679	1,306	52.3	$6.00
36	Hy4 previewTencent	61.4	$0.1212	87.5%	0.0%	4,190	2,100	844	1,246	44.3	$2.501
37	GLM 5.3 FlashZ.ai	61.4	$0.0058	100.0%	0.0%	4,184	2,169	717	1,298	57.7	$0.50
38	GPT-5.5OpenAI	59.7	$2.7357	87.5%	12.5%	3,692	2,189	628	875	67.9	$30.00
39	Muse Spark OpenMeta	59.2	$0.1216	100.0%	0.0%	4,112	2,089	657	1,366	47.3	$4.25
40	MiMo-V2.6-FlashXiaomi	58.4	$0.0178	62.5%	0.0%	4,180	2,179	430	1,571	54.8	$0.28
41	Muse Spark 1.2Meta	57.4	$0.1434	100.0%	0.0%	4,176	2,012	706	1,458	50.9	$4.25
42	Mistral Large 4Mistral AI	56.1	$0.1911	87.5%	0.0%	4,192	1,798	1,008	1,386	63.4	$2.10
43	GPT-6 LunaOpenAI	55.8	$0.0117	100.0%	0.0%	4,184	1,995	574	1,615	57.4	$0.50
44	Claude Haiku 5.5Anthropic	55.7	$0.0119	100.0%	0.0%	4,178	1,788	954	1,436	70.8	$0.50
45	Claude Sonnet 5Anthropic	54.4	$0.351	100.0%	0.0%	4,178	1,820	842	1,516	54.7	$10.00
46	GPT-5.6 LunaOpenAI	54.0	$0.0308	100.0%	0.0%	4,182	1,707	1,029	1,446	55.3	$1.20
47	DeepSeek V4 Pro 0813DeepSeek	53.9	$0.1466	75.0%	0.0%	4,192	1,851	737	1,604	56.1	$1.98
48	R Beam 1.0 501B A23BReflection	53.4	$0.0285	62.5%	0.0%	4,188	1,811	737	1,640	Not benchmarked	$1.20
49	GPT-5.6 TerraOpenAI	53.1	$0.3834	75.0%	0.0%	4,188	1,709	919	1,560	63.8	$12.00
50	Step 5 PreviewStepFun	52.2	$0.1294	77.8%	0.0%	2,935	1,264	395	1,276	61.2	$2.70
51	DeepSeek V4 Pro 0423DeepSeek	51.6	$0.0466	50.0%	0.0%	4,192	1,721	832	1,639	38.7	$3.1275
52	GLM 5.2Z.ai	49.9	$0.1219	75.0%	0.0%	4,182	1,464	1,160	1,558	61.1	$4.40
53	Kimi K3Moonshot AI	49.8	$0.076	87.5%	0.0%	4,184	1,545	987	1,652	62.3	$15.00
54	Claude Opus 4.5Anthropic	49.8	$0.5887	100.0%	0.0%	4,188	1,576	908	1,704	44.0	$25.00
55	Nemotron 3.5 Super VLNVIDIA	48.1	$0.0117	62.5%	0.0%	4,196	1,474	984	1,738	19.1	$0.50
56	Nemotron 3 UltraNVIDIA	47.7	$0.0181	75.0%	0.0%	1,180	472	260	448	30.0	$2.40
57	GPT-5OpenAI	46.7	$0.282	75.0%	0.0%	4,184	1,382	1,055	1,747	38.8	$10.00
58	GPT-5.4 MiniOpenAI	46.3	$0.2519	87.5%	0.0%	4,190	1,404	958	1,828	35.8	$4.50
59	Qwen3.8 27BQwen	43.6	$0.1957	62.5%	25.0%	3,176	1,705	235	1,236	74.1	$2.55
60	Ling 3.1 FlashinclusionAI	43.5	$0.0077	37.5%	0.0%	2,042	502	658	882	Not benchmarked	$0.225
61	DeepSeek V4 Flash 0423DeepSeek	43.0	$0.005	62.5%	0.0%	4,184	1,247	985	1,952	Not benchmarked	$0.28
62	Seed-2.0-CodeByteDance Seed	41.4	$0.0507	62.5%	0.0%	4,186	1,119	1,135	1,932	Not benchmarked	$3.00
63	GPT-4.1OpenAI	41.1	$0.0572	62.5%	0.0%	4,194	1,163	999	2,032	Not benchmarked	$8.00
64	MiniMax M3.1 FlashMiniMax	40.8	$0.00	62.5%	12.5%	3,696	1,429	519	1,748	58.9	$0.60
65	o3OpenAI	40.5	$0.0431	87.5%	0.0%	4,182	1,077	1,110	1,995	Not benchmarked	$8.00
66	Qwen3.7 MaxQwen	39.0	$0.0911	12.5%	12.5%	3,676	1,176	862	1,638	54.1	$4.425
67	Gemini 3.5 Flash LiteGoogle	37.3	$0.017	56.3%	6.3%	7,852	2,107	1,810	3,935	48.9	$2.50
68	Mercury 3Inception	37.3	$0.0028	87.5%	0.0%	4,178	991	1,007	2,180	29.3	$0.35
69	MiMo-V2.5-ProXiaomi	36.9	$0.0078	75.0%	0.0%	2,654	637	550	1,467	31.6	$0.87
70	Gemma 4 31BGoogle	36.3	$0.0038	62.5%	0.0%	4,190	831	1,254	2,105	39.0	$0.40
71	Kimi K2.7 CodeMoonshot AI	34.1	$0.0994	37.5%	37.5%	914	347	318	249	53.4	$4.00
72	Muse Glimmer 30BMeta	34.0	$0.0197	62.5%	12.5%	3,672	833	1,090	1,749	Not benchmarked	$1.20
73	gpt-oss-120bOpenAI	33.8	$0.0074	62.5%	12.5%	3,650	909	893	1,848	Not benchmarked	$0.50
74	T Inkling SmallThinking Machines	33.6	$0.0371	66.7%	11.1%	4,184	916	1,066	2,202	27.3	$1.20
75	Grok Build 0.1xAI	33.4	$0.0607	0.0%	0.0%	4,200	805	1,073	2,322	Not benchmarked	$2.00
76	Qwen3.7 PlusQwen	31.3	$0.0544	12.5%	25.0%	3,170	962	637	1,571	33.2	$1.28
77	T InklingThinking Machines	31.2	$0.2246	56.3%	12.5%	7,340	1,596	1,789	3,955	31.6	$4.05
78	Hy3 previewTencent	31.1	$0.0032	55.6%	22.2%	3,672	988	866	1,818	Not benchmarked	$0.60
79	GPT-5.4 NanoOpenAI	31.0	$0.0247	31.3%	25.0%	6,332	2,116	875	3,341	44.8	$1.25
80	P Laguna S 2.1RetiredPoolside	30.5	$0.003	50.0%	14.3%	6,330	1,849	1,090	3,391	Not benchmarked	$0.18
81	Step 3.7 FlashStepFun	29.5	$0.0142	37.5%	25.0%	3,094	673	976	1,445	Not benchmarked	$1.15
82	MiniMax M3MiniMax	29.0	$0.0254	50.0%	25.0%	3,166	896	567	1,703	33.6	$1.20
83	LongCat 2.0Meituan	27.6	$0.0134	50.0%	12.5%	7,370	1,510	1,392	4,468	Not benchmarked	$1.20
84	Mistral Medium 3.5Mistral AI	27.0	$0.1747	37.5%	12.5%	3,650	593	948	2,109	25.4	$7.50
85	MiMo-V2.5Xiaomi	23.8	$0.0024	43.8%	31.3%	5,768	1,269	1,311	3,188	50.0	$2.00
86	Gemini 3.1 Flash LiteGoogle	19.2	$0.0101	25.0%	25.0%	3,184	453	658	2,073	Not benchmarked	$1.50
87	Ling 3.0 FlashinclusionAI	18.1	$0.00	37.5%	37.5%	5,202	712	1,467	3,023	29.6	$0.1215

Only models in our catalog are listed, and each appears once with its best published run. 87 rows were read from the source on November 5, 2026, and rows that match no catalog model are left out. Cost per program is what DuelLab reports the model spent, on average, to write one player program, in US dollars. It covers writing the code, not playing the matches. Every other column is a figure the board publishes for the same run.

Data: DuelLab GameBench 2, licensed CC BY 4.0.

What the result is made of

GameBench 2: the parts of each result
Model	Result	Go	Chess	Shogi	Xiangqi	Hex	International Draughts	Game of the Amazons	Tumbleweed
Anthropic Claude Opus 5.5	94.9	82.7	95.7	97.3	94.8	98.3	99.4	97.5	93.1
Anthropic Claude Fable 5.1	94.1	92.1	95.3	95.3	89.4	92.5	97.8	93.1	97.5
OpenAI GPT-6 Astra	94.0	96.9	95.3	97.8	90.7	99.0	84.9	97.3	90.3
OpenAI GPT-6.1 Sol	91.9	74.5	94.3	97.8	94.8	98.0	87.7	95.5	93.0
Anthropic Claude Sonnet 5.5	91.3	78.6	96.0	94.8	82.3	97.2	90.1	92.2	98.9
Pareto 26.10	89.7	90.5	78.6	95.9	85.3	90.7	91.8	94.7	90.0
Google Gemini 4 Argon	89.0	94.7	94.4	91.6	70.7	95.4	96.9	82.3	86.3
OpenAI GPT-6.1 Terra	85.7	73.4	76.8	88.3	83.6	95.1	89.9	97.0	81.4
OpenAI GPT-6.1 Terra Pro	84.8	40.1	88.3	94.7	83.7	95.8	90.6	95.5	90.0
Anthropic Claude Opus 5	84.7	68.5	82.1	93.3	79.5	92.0	Not published	95.3	82.4
MoonshotAI Kimi K3.1 Pro	84.0	91.9	78.8	85.7	82.6	88.7	82.2	72.3	89.5
OpenAI GPT-5.6 Sol	79.6	73.9	75.3	77.2	82.1	73.5	82.8	86.7	85.2
OpenAI GPT-6 Sol	79.3	9.7	82.1	93.1	87.4	93.1	88.1	93.4	87.5
Meta Muse Wonder	77.1	83.8	66.0	81.4	80.4	92.0	83.4	71.9	58.2
Anthropic Claude Fable 5	75.0	84.4	79.5	84.8	84.4	53.7	84.9	63.8	64.2
Z.ai GLM 5.3	72.1	0.0	72.2	82.8	76.0	88.2	89.3	76.4	91.8
SpaceXAI Grok 4.7	72.0	78.0	79.3	69.8	92.3	92.1	0.0	74.2	90.5
Xiaomi MiMo-V2.6-Pro	71.3	51.9	79.0	83.9	45.6	83.2	54.2	87.9	84.8
DeepSeek V4.1 Flash 1017	69.7	67.3	79.6	45.6	69.2	70.7	74.7	70.5	80.3
Meta Muse Spark 1.3	68.7	75.0	60.9	80.9	50.1	76.4	82.2	56.6	67.5
SpaceXAI Grok 4.6	68.6	59.8	53.6	65.0	64.3	79.7	75.7	74.3	76.1
Qwen3.8 Max (0902)	68.1	74.6	48.1	69.7	83.1	90.9	61.0	74.4	43.3
Z.ai GLM 5.3 FlashX	67.6	40.6	67.8	82.6	77.1	75.8	79.8	70.4	46.6
DeepSeek V4.1 Flash	67.4	60.2	56.2	58.3	73.6	79.6	79.2	73.5	58.7
Qwen3.8 2.4T A95B	66.7	57.1	52.5	79.3	45.6	65.9	77.4	67.5	88.4
Google Gemini 3.1 Pro Preview	64.7	72.4	42.2	64.0	80.5	82.5	17.9	72.9	85.0
Google Gemini 3.8 Flash	64.2	53.4	69.0	69.7	46.5	63.8	86.4	67.8	56.9
Elorian Glimpse	64.1	82.0	43.0	35.5	54.3	55.6	81.7	73.1	87.4
Fireworks Ember-1	62.9	74.9	41.5	79.1	39.1	60.3	49.0	73.7	85.7
Google Gemini 3.6 Flash	62.8	54.5	54.1	66.6	62.7	81.1	79.4	49.3	54.7
Google Gemini 3.5 Flash	62.6	89.9	61.5	69.0	41.2	48.8	71.0	50.6	69.0
Anthropic Claude Opus 4.8	62.4	67.0	78.3	0.0	70.0	46.2	79.7	74.5	83.2
OpenAI GPT-5.4	61.8	94.8	0.0	42.3	55.1	84.0	78.6	75.5	64.2
Qwen3.8 Flash	61.7	50.4	57.0	65.0	47.9	68.5	78.1	73.3	53.1
SpaceXAI Grok 4.5	61.4	82.1	57.7	43.3	43.1	71.0	51.1	54.5	88.7
Tencent Hy4 preview	61.4	52.9	77.1	70.1	54.9	67.4	77.9	65.7	25.0
Z.ai GLM 5.3 Flash	61.4	69.2	45.3	78.3	73.0	45.6	78.5	73.7	27.4
OpenAI GPT-5.5	59.7	79.5	0.0	49.0	36.4	75.3	77.7	85.1	74.8
Meta Muse Spark Open	59.2	90.9	44.4	65.3	38.0	38.1	47.2	71.9	77.9
Xiaomi MiMo-V2.6-Flash	58.4	50.2	65.5	60.5	57.7	56.9	62.8	81.1	32.9
Meta Muse Spark 1.2	57.4	78.8	46.9	59.0	47.8	58.0	25.1	72.7	71.0
Mistral Large 4	56.1	33.1	64.8	42.2	66.0	46.5	76.4	77.3	42.2
OpenAI GPT-6 Luna	55.8	28.5	62.4	48.6	63.3	50.6	83.0	38.5	71.5
Anthropic Claude Haiku 5.5	55.7	44.0	61.5	65.2	68.1	54.3	65.2	55.2	31.9
Anthropic Claude Sonnet 5	54.4	61.0	41.1	59.1	51.5	51.2	26.1	66.2	79.1
OpenAI GPT-5.6 Luna	54.0	79.4	43.9	40.2	37.8	48.4	80.7	70.6	30.8
DeepSeek V4 Pro 0813	53.9	41.6	43.9	61.9	41.8	71.3	62.5	77.3	30.8
Reflection Beam 1.0 501B A23B	53.4	23.8	49.0	76.0	59.7	50.1	71.6	38.1	59.1
OpenAI GPT-5.6 Terra	53.1	68.6	58.1	35.3	41.3	52.6	77.1	42.9	48.4
StepFun Step 5 Preview	52.2	62.5	64.7	51.7	36.1	59.9	66.4	54.6	21.4
DeepSeek V4 Pro 0423	51.6	60.5	38.9	28.0	43.0	46.0	45.9	69.3	81.0
Z.ai GLM 5.2	49.9	45.0	52.2	26.9	55.0	49.6	82.7	63.8	23.7
MoonshotAI Kimi K3	49.8	49.9	41.3	47.4	44.3	51.9	44.8	28.0	90.5
Anthropic Claude Opus 4.5	49.8	70.2	43.0	48.8	41.0	48.5	52.8	48.6	45.2
NVIDIA Nemotron 3.5 Super VL	48.1	64.0	43.7	34.7	40.0	55.2	43.5	55.4	47.9
NVIDIA Nemotron 3 Ultra	47.7	54.2	41.8	52.5	44.6	84.2	35.3	37.6	31.3
OpenAI GPT-5	46.7	37.6	41.3	35.8	38.4	72.7	66.2	51.2	30.7
OpenAI GPT-5.4 Mini	46.3	68.6	40.8	42.8	43.0	45.3	44.7	44.7	40.9
Qwen3.8 27B	43.6	72.6	0.0	0.0	69.1	56.2	26.4	76.1	48.4
inclusionAI Ling 3.1 Flash	43.5	46.9	43.9	38.9	43.0	59.1	24.6	23.2	68.6
DeepSeek V4 Flash 0423	43.0	31.0	40.9	45.5	39.4	47.3	46.2	35.4	58.3
ByteDance Seed Seed-2.0-Code	41.4	39.7	40.4	40.1	37.3	46.8	18.9	44.0	64.1
OpenAI GPT-4.1	41.1	59.2	42.1	37.4	37.4	12.6	46.9	46.5	47.1
MiniMax M3.1 Flash	40.8	66.2	0.0	30.8	37.8	47.7	48.3	69.1	26.8
OpenAI o3	40.5	36.2	42.1	39.9	43.1	48.7	44.7	38.1	31.6
Qwen3.7 Max	39.0	17.7	43.4	0.0	42.8	36.9	57.5	74.1	39.8
Google Gemini 3.5 Flash Lite	37.3	34.7	39.5	26.5	46.7	10.8	53.2	60.1	26.6
Inception Mercury 3	37.3	33.1	41.7	32.5	39.9	23.0	63.1	32.8	32.0
Xiaomi MiMo-V2.5-Pro	36.9	54.7	41.4	53.1	38.5	19.5	24.9	39.1	24.4
Google Gemma 4 31B	36.3	20.9	42.6	41.3	42.0	69.0	25.2	17.5	31.7
MoonshotAI Kimi K2.7 Code	34.1	63.4	43.3	0.0	38.9	0.0	48.6	0.0	78.5
Meta Muse Glimmer 30B	34.0	29.9	42.3	36.8	48.1	43.2	0.0	46.3	25.7
OpenAI gpt-oss-120b	33.8	0.0	42.8	44.7	41.1	55.2	21.8	33.0	32.1
Thinking Machines Inkling Small	33.6	31.2	41.1	43.8	20.7	39.5	51.0	18.3	22.9
SpaceXAI Grok Build 0.1	33.4	35.6	43.0	28.0	39.5	48.1	30.9	14.7	27.2
Qwen3.7 Plus	31.3	49.6	42.1	0.0	0.0	42.0	66.2	23.7	26.3
Thinking Machines Inkling	31.2	38.0	40.9	46.5	21.4	15.0	25.7	28.0	33.7
Tencent Hy3 preview	31.1	63.1	38.0	0.0	42.6	28.6	12.6	25.1	38.9
OpenAI GPT-5.4 Nano	31.0	15.2	19.5	0.0	25.1	48.4	56.5	34.2	49.1
Poolside Laguna S 2.1Retired	30.5	48.3	20.0	38.2	0.0	46.5	18.9	43.9	27.9
StepFun Step 3.7 Flash	29.5	30.9	42.6	35.9	39.5	0.0	46.4	40.5	0.0
MiniMax M3	29.0	56.5	39.3	38.3	0.0	27.2	0.0	22.4	48.6
Meituan LongCat 2.0	27.6	41.2	19.8	33.7	22.6	11.2	33.6	34.5	24.1
Mistral Medium 3.5	27.0	0.0	44.1	35.9	38.6	34.7	23.8	13.0	25.9
Xiaomi MiMo-V2.5	23.8	44.5	20.3	0.0	40.7	15.0	37.7	14.2	17.8
Google Gemini 3.1 Flash Lite	19.2	32.3	0.0	0.0	41.6	16.2	20.9	6.2	36.7
inclusionAI Ling 3.0 Flash	18.1	33.5	19.8	28.0	21.9	0.0	8.5	0.0	33.2

The parts as DuelLab GameBench 2 publishes them, for the same run as the result.

Best model per game

GameBench 2: the leading model of each game
Game	Leader	Score	Median across models
Go	OpenAI GPT-6 Astra	96.9	57.1
Chess	Anthropic Claude Sonnet 5.5	96.0	44.1
Shogi	OpenAI GPT-6 Astra	97.8	48.8
Xiangqi	OpenAI GPT-6.1 Sol	94.8	45.6
Hex	OpenAI GPT-6 Astra	99.0	55.2
International Draughts	Anthropic Claude Opus 5.5	99.4	62.6
Game of the Amazons	Anthropic Claude Opus 5.5	97.5	65.7
Tumbleweed	Anthropic Claude Sonnet 5.5	98.9	54.7

How a run works

DuelLab gives the model the rules of one game and the interface its player must use. The model answers with the source code of a complete player program.
The code is built and run in a sandbox with fixed limits on time, memory, CPU, files and network. A program that fails can be sent back for repair; one that never works counts as 0 for that game. A program too slow for the speed check is left out rather than scored.
Every working program plays every other program in the game's pool, once with each side. Crashes and illegal moves count as losses.
Wins, draws and losses go into a Bradley–Terry rating. DuelLab turns the rating into a score from 0 to 100: the share of points a program would be expected to take from a fixed reference field of programs.
The headline score is the average of the eight game scores. Each model and reasoning setting is ranked as its own row; our model pages show a model's best setting and list the others as variants.

At a glance

Run by: DuelLab
Task: Write a program that plays a strategy game; the program then plays
Games: 8 (Go, Chess, Shogi, Xiangqi, Hex, International Draughts, Game of the Amazons, Tumbleweed)
Score: 0 to 100, relative to a fixed reference field
Size of the 26 October 2026 release: 525,020 rated matches between 2,053 player programs
Releases: 21 between July and October 2026
Data licence: CC BY 4.0 (CSV, JSON and Parquet)
In our indexes: No, shown beside the coding rankings

Role in the index

A coding board we show beside the Vibe Coding Index, not inside it: the index keeps the six boards it was defined on, so a score does not move every time a board is added. Read the methodology

Questions

Is a GameBench 2 score of 95 the share of games won?: No. It is the share of points a model's programs would be expected to take from DuelLab's fixed reference field of programs. A model at 95 wins nearly every match against that field. DuelLab refits the ratings at every release, so a score can move without the model changing.
Why does one model have several GameBench 2 scores?: DuelLab runs each model at several reasoning settings, such as Low, Medium and XHigh, and ranks each one separately. We show a model's best setting as its result and list the other settings as variants on its page.
What do first-try builds and failed programs mean?: First-try builds is the share of a model's programs that built and passed DuelLab's checks without a repair round. Failed programs is the share that never worked and scored 0. A repair that works is not penalised, so a model can score well with a lower first-try rate.
How is GameBench 2 different from chess leaderboards where models play?: On boards like Kaggle Game Arena the model chooses every move itself. In GameBench 2 the model never plays: it writes a program once, and the program plays hundreds of matches. It measures whether a model can turn the rules of a game into a working, strong engine.
Why is GameBench 2 not in the LLMBenchmark Index?: Its score is relative to DuelLab's reference field and is refitted at every release. Adding it would move our index each time DuelLab publishes, so we show it beside the coding rankings instead.