Benchmarko1 previewClaude 3.5 SonnetLlama 3.1 405BGPT 4oo1 miniQwen2.5 72BClaude 3 OpusGemini 1.5 ProDeepSeek V2.5Llama 3.1 70BMistral Large 2GPT 4o miniGemini 1.5 FlashLlama 3.1 8BGLM 4 PlusYi LargeDeepSeek V4.1 FlashQwen3.8 Flash NextGLM 5.3GLM 5.3 FlashDeepSeek V4 Pro 0813DeepSeek V4 Flash 0731MiMo V2.6 ProMiMo V2.6 FlashMiMo V2.5 ProKimi K3Claude Opus 5GPT 5.6 SolClaude Fable 5GPT 6 AstraClaude Fable 5.1Qwen3.8 MaxQwen3.8 27BQwen3.6 PlusQwen3.6 27BQwen3.6 35B A3BClaude Opus 5.5
Agents' Last Exam%————————————————31.8—28.5—25.725.231.627.613.227.628.626.725.734.2———————
Arena Elo12581268122512871230124912401255—11801224120712011140—1195—————————————————————
Automation Bench v1.0.6%————————————————54.8—48.8—43.237.753.152.316.046.750.345.846.252.0———————
BabyVision (with tools)%————————————————89.6————————85.794.188.9—————————
BBH%—93.193.083.1—86.893.789.2—81.678.975.075.062.3———————————————————————
Chartography (with tools)%————————————————78.9————————68.184.079.9——88.4—————89.0
Codeforces (Rating)————————————————3471———33483289———————————————
CursorBench 4.0%——————————————————————————46.641.7——51.8—————57.8
CyberGym%————————————————88.1—84.5—83.376.794.095.140.080.0—84.5—————————
DeepSWE v1.1%————————————————74.258.766.963.462.754.471.967.919.067.574.073.070.074.0—56.642.2—13.3——
ExploitBench%——————————————————————47.925.316.632.270.078.578.0100.083.0——————
ExploitGym%————————————————15.3—15.0—5.41.817.86.00.18.122.133.728.442.430.4——————
FrontierCode v1.1 (Main)%——————————————————————————48.047.5—53.350.3—————54.4
GDPVal 2.1 (AA)————————————————1600—————1673—1107152417081588159515421735—————1846
GPQA Diamond%78.365.051.153.660.049.050.446.248.041.748.040.238.630.4——90.991.788.1—92.489.9———92.993.494.1———92.689.290.487.886.0—
GSM8K%—96.496.895.8—95.895.091.791.095.193.087.086.284.5—89.0—————————————————————
HLE%————————————————36.835.9———————43.556.344.5———43.630.828.824.021.4—
HLE (with tools)%————————————————63.9—62.555.360.051.5———59.863.6——57.265.656.2—50.6——67.7
HumanEval%92.492.089.090.292.486.684.984.189.080.592.087.274.372.681.080.0—————————————————————
IFEval%—88.088.684.0—84.085.0—84.687.5—80.5—80.0———————————————————94.3———
JobBench%————————————————45.855.7————62.061.225.054.365.745.457.4——53.433.4—21.8——
LiveCodeBench v6%—————————————————91.9——————————————90.387.183.980.4—
MATH%94.871.173.876.690.083.160.167.772.767.968.670.254.951.9———————————————————————
MathArena Apex%————————————————65.6———65.358.6———65.6———————————
MMLU%90.888.787.388.785.286.186.885.980.683.684.082.078.969.486.084.6—————————————————————
MMLU-Pro%79.278.073.372.674.071.168.075.869.466.461.258.567.348.368.0——————————————————88.586.285.2—
NL2Repo%————————————————65.448.158.056.361.554.2———58.075.356.8———55.942.337.936.229.4—
OSWorld 2.0 (partial)%——————————————————————————74.0———80.7—————81.8
OSWorld-Verified%——————————————————————82.080.8—84.883.483.086.0———84.362.563.9——
ProgramBench%————————————————20.3—19.0—15.5—26.526.012.517.537.023.033.0————————
SEC Bench Pro%————————————————62.8———56.430.966.347.517.7——74.3—85.4———————
SWE-bench Pro%—————————————————62.5—————————————67.761.756.653.549.5—
Terminal Bench 2.1%————————————————90.6—88.284.387.982.789.987.665.288.389.188.884.389.991.486.673.0————
Terminal Bench 3.0%————————————————30.0—28.3—11.87.6———17.743.334.4—————————
Terminal Bench 4.0%————————————————31.2—37.9—12.47.034.928.81.512.651.839.942.457.955.8—————66.4
Terminal-Bench-Science 0.1%——————————————————————————29.022.4—64.652.6—————58.7
Toolathlon-verified%——————————————————————76.973.649.176.580.674.977.9—77.8——————
ZeroBench-main (with tools)%————————————————49.0————————41.052.053.0—————————

Figures are taken from public model cards, technical reports and vendor announcements, and are approximate where sources disagree. Results from different benchmark versions or tool-use settings never share a column: Terminal Bench 2.1, 3.0 and DeepSWE v1.1 are distinct benchmarks, and HLE with tools is separate from HLE without tools. Pick up to 5 models above — type to search by name or provider. Benchmarks none of them published are left out, and the best result in each row is shown in bold.