AI chess accuracy

Kimi K3.1 Pro (Max) is coming to chessbench-ai.github.io.

ChessBench now ranks all 41 models by Accuracy instead of benchmark Elo. The new system puts engine-evaluated move quality at the center of the leaderboard: higher Accuracy means less loss of estimated winning chances across the evaluated moves.

The available game samples were reanalyzed with Stockfish 19 at depth 12. Scores use the Lichess Accuracy formula, with model results weighted by the number of evaluated moves. Rankings use the full scores, while the leaderboard displays two decimal places.

Elo is no longer the primary ranking metric. It remains available alongside ACPL in each model's details. Accuracy describes the recorded samples, not a human chess rating; sample size and incomplete games still matter when comparing results.

Original update: Gemini 4 Argon (High): 2,430 Elo at $0.45 per task

Gemini 4 Argon (High) joins the 41-model ChessBench field at rank 12 with 2,430 benchmark Elo, an average centipawn loss (ACPL) of 12.27, and a reported cost of $0.45/task. It sits between GLM-5.3 Flash at 2,450 Elo and Claude Opus 5 (high) at 2,400 Elo.

Compared with Gemini 3.1 Pro, the new entry scores 190 Elo higher and reduces ACPL from 34.14 to 12.27 — a decrease of 21.87 centipawns, or approximately 64.1%. Gemini 3.1 Pro (High) remains 140 Elo ahead at 2,570, with a lower ACPL of 5.96.

The $0.45/task figure is the reported cost for this evaluation. Benchmark Elo measures performance relative to the tested ChessBench field.

Original update: Claude Sonnet 5.5 (Medium): 2,350 Elo at $0.16 per task

Claude Sonnet 5.5 (Medium) joins the 40-model ChessBench field at rank 15 with 2,350 benchmark Elo, an average centipawn loss (ACPL) of 19.75, and a reported cost of $0.16/task. It sits between Claude Fable 5.1 (Max) at 2,370 Elo and GPT-6 Pro (Max) at 2,340 Elo.

Compared with Claude Sonnet 5 (Max), the new entry scores 670 Elo higher and reduces ACPL from 70.02 to 19.75 — a decrease of 50.27 centipawns, or approximately 71.8%. Claude Sonnet 5 (Max) now appears at rank 34. Its cost has not been reported, so a direct cost comparison is not available.

Compared with Claude Opus 5.5 (Medium), it scores 70 Elo higher and reduces ACPL from 31.18 to 19.75 — a decrease of 11.43 centipawns, or approximately 36.7%. The reported cost falls from $0.58/task to $0.16/task, a decrease of $0.42. Opus 5.5 (Medium) costs approximately 3.63 times as much. Claude Fable 5.1 (Max) remains 20 Elo ahead with an ACPL of 18.54, at a reported cost of $44.46/task — approximately 278 times the cost of this entry.

These comparisons describe the recorded ChessBench results, not performance outside the benchmark. Costs are the reported figures for these evaluations, not general model pricing. Benchmark Elo is relative to the tested field and is not a human chess rating.

Original update: GPT-6.1 Sol (Max): 2,550 Elo at $0.31 per task

GPT-6.1 Sol (Max) joins the 39-model ChessBench field at rank 7 with 2,550 benchmark Elo, an average centipawn loss (ACPL) of 6.54, and a reported cost of $0.31/task. It sits between Gemini 3.1 Pro (High) at 2,570 Elo and Qwen 3.8 Max at 2,500 Elo.

Compared with GPT-6 Sol (Max), the new entry scores 60 Elo higher and reduces ACPL from 9.45 to 6.54 — a decrease of 2.91 centipawns, or approximately 30.8%. GPT-6 Sol (Max) now appears at rank 9. The new result remains 110 Elo behind the leaderboard’s top entry, GPT-5.6 Sol Pro (Max), at 2,660 Elo.

The higher score comes at a higher reported cost: $0.31/task versus $0.14/task for GPT-6 Sol (Max), an increase of $0.17. That is approximately 2.21 times as much. Grok 4.7 (xhigh) remains ahead on all three reported metrics, with 2,600 Elo, 5.52 ACPL, and $0.21/task.

These comparisons describe the recorded ChessBench results, not performance outside the benchmark. Costs are the reported figures for these evaluations, not general model pricing. Benchmark Elo is relative to the tested field and is not a human chess rating.

Claude Opus 5.5 (Medium) joins the 38-model ChessBench field with 2,280 benchmark Elo, an average centipawn loss (ACPL) of 31.18, and a reported cost of $0.58/task. It is listed at rank 19, level on Elo with GPT-4.5 at rank 18.

Among the adaptive-thinking Claude entries, it scores 20 Elo above Opus 4.7 and 30 Elo above Opus 4.8. Compared with Opus 4.8, its ACPL is 0.76 lower (31.18 versus 31.94). Opus 4.7 has a lower ACPL of 26.78 despite its lower Elo; the two metrics capture different aspects of the recorded games.

The $0.58/task figure is the reported cost for this evaluation. Benchmark Elo compares models within this ChessBench field and is not a human chess rating.

Two new results bring the ChessBench leaderboard to 37 models. GPT-6 Sol (Max) enters at rank 8 with 2,490 benchmark Elo, an average centipawn loss (ACPL) of 9.45, and a reported cost of $0.14/task. Space Bunny Alpha enters at rank 34 with 1,550 benchmark Elo, 81.24 ACPL, and a reported cost of $0.00/task.

GPT-6 Sol (Max) sits between Qwen 3.8 Max at 2,500 Elo and Grok 4.6 (xhigh) at 2,480 Elo. Compared with the existing GPT-6 Pro (Max) entry, it scores 150 Elo higher and reduces ACPL from 23.45 to 9.45 — approximately 59.7% lower. It remains 170 Elo behind the current leader, GPT-5.6 Sol Pro (Max), at 2,660 Elo. These comparisons describe the recorded benchmark results, not a claim about model lineage or performance outside ChessBench.

Space Bunny Alpha sits between Gemini 3.5 Flash at 1,650 Elo and GPT-OSS-120b at 1,460 Elo. Its $0.00/task result is the lowest reported cost in the current dataset, while its benchmark Elo places it near the bottom of the field. Separately, OpenRouter’s model listing identifies it as an anonymous stealth model, so its provider is listed as Anonymous rather than attributed to an unconfirmed developer.

GPT-6 Sol (Max) is reported at $0.14/task, while Space Bunny Alpha is reported at $0.00/task for this evaluation. Space Bunny Alpha’s $0.00 is the reported cost for this evaluation, not a guarantee of future pricing. As always, benchmark Elo is relative to the ChessBench field and is not a human chess rating.

Original update: Grok 4.7 (xhigh): 2,600 Elo at $0.21 per task

Grok 4.7 (xhigh) joins ChessBench with 2,600 benchmark Elo, an average centipawn loss (ACPL) of 5.52, and a reported cost of $0.21/task. It enters at rank 4 out of 35 models, between GPT-5.5 Pro at 2,630 Elo and Kimi K3 (Max) at 2,590 Elo.

Compared with Grok 4.6 (xhigh), the new result is 120 Elo higher, while average ACPL falls from 9.59 to 5.52 — a reduction of approximately 42.4%. This makes Grok 4.7 (xhigh) the highest-rated xAI entry in the current ChessBench field.

The reported cost is $0.21 per task. A direct cost comparison with Grok 4.6 is not available because its cost has not been reported in the dataset. These results describe performance within ChessBench; benchmark Elo is relative to the tested field, not a human chess rating.

Pareto 26.9 (Union Alpha) has been added with 2,330 benchmark Elo, 25.01 ACPL, and a cost of $0.05/task. It enters at rank 13, between GPT-6 Pro (Max) and Gemini 3.6 Flash.

Gemini 3.8 Flash (medium) has been added with 2,300 benchmark Elo, 30.74 ACPL, and a cost of $0.08/task. It enters at rank 13, after Gemini 3.6 Flash, which also has 2,300 benchmark Elo, and ahead of GPT-4.5.

Claude Fable 5.1 (Max) has been added as a separate model with 2,370 benchmark Elo, 18.54 ACPL, and a cost of $44.46/task. It enters at rank 10, between Gemma 4 and GPT-6 Pro (Max).

The leaderboard now includes Cost per Task. Costs not yet reported are shown as a dash (—). ACPL and cost are also available in model details.

GPT-6 Pro (Max) has been added to ChessBench with a benchmark Elo of 2,340. It enters at rank 10, between Gemma 4 and Gemini 3.6 Flash.

The leaderboard now includes 30 models across seven providers.

Gemini 3.7 Flash (High) has been added to ChessBench with a benchmark ELO of 2,230 It enters at rank 17, between MiniMax M3 and ChatGPT 5.4 Pro.

The leaderboard consolidates GPT-5.6 Sol Pro (Max) and its Stealth Testing checkpoint into a single rank-one entry, with the checkpoint’s 2,650 ELO shown as secondary information. Claude Fable 5 (Max) therefore moves to rank 2.

Grok 4.6 (xhigh) has been added to ChessBench with a benchmark ELO of 2,480 The result places it at rank 7, between Qwen 3.8 Max and Claude Opus 5 (high), and makes it the strongest xAI model in the current field.

Qwen 3.8 Max joins with 2,500 benchmark ELO. It places at rank 6, between Gemini 3.1 Pro (High) and Grok 4.6 (xhigh), and adds Alibaba to the provider field.

Claude Opus 5 (high) enters at 2,400 benchmark ELO. Compared to the original Opus 5 configuration, this represents a substantial improvement in playing strength, moving the entry closer to the current frontier.

Gemini 3.6 Flash joins at 2,300 benchmark ELO, alongside Gemini 3.5 Flash-Lite at 2,200 benchmark ELO. Flash-Lite’s benchmark ELO is lower because illegal moves and several quick terminations affected game outcomes.

Grok 4.5 was rebenchmarked after errors were found in the original evaluation. Its benchmark ELO changes from 1,457 to 1,680.

Claude Fable 5 (Max) from Anthropic and Kimi K3 (Max) from Moonshot AI join the leading group. Claude Fable 5 (Max) holds rank 2 with 2,640 benchmark ELO, while Kimi K3 (Max) holds rank 4 with 2,590 benchmark ELO.

The additions broaden the provider mix near the top of the table and add two high-strength reference points to the leaderboard.

ChessBench TV replays complete model-vs-model games on an interactive board, with the full move score and playback controls. The first replay features GPT-5.6 Sol Pro against Gemini 3.1 Pro.