aiapicostcalcUnit economicsPricesBrandsCompareBenchmarksSkills & PluginsFind my modelGuidesChanges

LLM benchmark comparison: price vs performance

Short answer: There is no single best LLM here: every benchmark measures a different skill. The highest LMArena Elo score is Claude Opus 5.5 at 1,526; one typical agent task on it costs $0.1168. The cheapest verified model per typical agent task is Gemini 2.5 Flash-Lite at $0.00277. Each score below keeps its own table, source and check date, and prices sit next to scores so you can see what one quality point costs.

Last verified 2026-11-18 · Sources: the benchmark leaderboards named in src/benchmarks.json and the prices in src/prices.json · Report an error

By Shangyi Duan · Benchmark sources and dates are listed per metric below

Leaderboards by metric

Sorted by each metric separately; prices are the same verified list rates across every table.

AIME

36 of 55 verified models report this metric · unit: % · Higher is better · source: leaderboard

#ModelScorePer agent taskScore checkedSource
1GPT-5.6 Sol91.4%$0.12072026-11-04source
2Claude Opus 590.9%$0.15092026-11-04source
3Claude Sonnet 586.9%$0.06042026-11-04source
4DeepSeek V4 Pro86.8%$0.03152026-11-04source
5Gemini 3.6 Flash86.6%$0.02262026-11-04source
6Gemini 2.5 Pro85%$0.04712026-11-04source
7GPT-5.6 Luna82.2%$0.006542026-11-04source
8Gemini 3.5 Flash-Lite80.9%$0.01162026-11-04source
9GPT-6 Astra80.8%$0.30182026-11-04source
9GPT-5.6 Terra80.8%$0.06542026-11-04source
9GPT-5.580.8%$0.16342026-11-04source
9GPT-5.480.8%$0.08172026-11-04source
9Claude Fable 5.180.8%$0.28732026-11-04source
9Claude Sonnet 4.680.8%$0.09052026-11-04source
15GPT-6.1 Sol78.2%$0.05842026-11-04source
15GPT-6.1 Terra78.2%$0.02922026-11-04source
15GPT-6 Sol78.2%$0.06042026-11-04source
15GPT-5.4 mini78.2%$0.02452026-11-04source
15GPT-4.178.2%$0.06112026-11-04source
15o4-mini78.2%$0.03362026-11-04source
15Claude Sonnet 5.578.2%$0.05842026-11-04source
15Claude Haiku 4.578.2%$0.03022026-11-04source
15Gemini 4 Argon78.2%$0.05842026-11-04source
15Gemini 3.8 Flash78.2%$0.02262026-11-04source
25GPT-5.4 nano75.1%$0.006662026-11-04source
25GPT-5 mini75.1%$0.009422026-11-04source
25Gemini 3.1 Flash-Lite75.1%$0.008172026-11-04source
25Gemini 2.5 Flash75.1%$0.01162026-11-04source
25Mistral Large 475.1%$0.01762026-11-04source
30Claude Opus 5.574.8%$0.11682026-11-04source
31GPT-6 Luna71.5%$0.003022026-11-04source
31Claude Haiku 5.571.5%$0.003022026-11-04source
31Gemini 2.5 Flash-Lite71.5%$0.002772026-11-04source
31Mistral Small 471.5%$0.006752026-11-04source
31Codestral 25.0871.5%$0.01272026-11-04source
36GPT-4o mini64.6%$0.005312026-11-04source

LiveCodeBench

36 of 55 verified models report this metric · unit: % · Higher is better · source: leaderboard

#ModelScorePer agent taskScore checkedSource
1GPT-5.6 Luna72.6%$0.006542026-11-04source
2GPT-6 Astra71.3%$0.30182026-11-04source
2GPT-5.6 Terra71.3%$0.06542026-11-04source
2GPT-5.571.3%$0.16342026-11-04source
2GPT-5.471.3%$0.08172026-11-04source
2Claude Fable 5.171.3%$0.28732026-11-04source
2Claude Sonnet 4.671.3%$0.09052026-11-04source
8Gemini 2.5 Pro70.9%$0.04712026-11-04source
9Claude Opus 569.7%$0.15092026-11-04source
10GPT-6.1 Sol69%$0.05842026-11-04source
10GPT-6.1 Terra69%$0.02922026-11-04source
10GPT-6 Sol69%$0.06042026-11-04source
10GPT-5.4 mini69%$0.02452026-11-04source
10GPT-4.169%$0.06112026-11-04source
10o4-mini69%$0.03362026-11-04source
10Claude Sonnet 5.569%$0.05842026-11-04source
10Claude Haiku 4.569%$0.03022026-11-04source
10Gemini 4 Argon69%$0.05842026-11-04source
10Gemini 3.8 Flash69%$0.02262026-11-04source
20GPT-5.6 Sol68.8%$0.12072026-11-04source
21Claude Sonnet 568.2%$0.06042026-11-04source
22Claude Opus 5.567.2%$0.11682026-11-04source
23DeepSeek V4 Pro66.4%$0.03152026-11-04source
24GPT-5.4 nano66.3%$0.006662026-11-04source
24GPT-5 mini66.3%$0.009422026-11-04source
24Gemini 3.1 Flash-Lite66.3%$0.008172026-11-04source
24Gemini 2.5 Flash66.3%$0.01162026-11-04source
24Mistral Large 466.3%$0.01762026-11-04source
29Gemini 3.6 Flash63.8%$0.02262026-11-04source
30GPT-6 Luna63.1%$0.003022026-11-04source
30Claude Haiku 5.563.1%$0.003022026-11-04source
30Gemini 2.5 Flash-Lite63.1%$0.002772026-11-04source
30Mistral Small 463.1%$0.006752026-11-04source
30Codestral 25.0863.1%$0.01272026-11-04source
35Gemini 3.5 Flash-Lite60.7%$0.01162026-11-04source
36GPT-4o mini47.6%$0.005312026-11-04source

SWE-bench Verified

36 of 55 verified models report this metric · unit: % · Higher is better · source: leaderboard

#ModelScorePer agent taskScore checkedSource
1GPT-5.6 Luna69%$0.006542026-11-04source
2GPT-6 Astra67.8%$0.30182026-11-04source
2GPT-5.6 Terra67.8%$0.06542026-11-04source
2GPT-5.567.8%$0.16342026-11-04source
2GPT-5.467.8%$0.08172026-11-04source
2Claude Fable 5.167.8%$0.28732026-11-04source
2Claude Sonnet 4.667.8%$0.09052026-11-04source
8Gemini 2.5 Pro67.5%$0.04712026-11-04source
9Claude Opus 566.3%$0.15092026-11-04source
10GPT-6.1 Sol65.6%$0.05842026-11-04source
10GPT-6.1 Terra65.6%$0.02922026-11-04source
10GPT-6 Sol65.6%$0.06042026-11-04source
10GPT-5.4 mini65.6%$0.02452026-11-04source
10GPT-4.165.6%$0.06112026-11-04source
10o4-mini65.6%$0.03362026-11-04source
10Claude Sonnet 5.565.6%$0.05842026-11-04source
10Claude Haiku 4.565.6%$0.03022026-11-04source
10Gemini 4 Argon65.6%$0.05842026-11-04source
10Gemini 3.8 Flash65.6%$0.02262026-11-04source
20GPT-5.6 Sol65.4%$0.12072026-11-04source
21Claude Sonnet 564.9%$0.06042026-11-04source
22Claude Opus 5.563.9%$0.11682026-11-04source
23DeepSeek V4 Pro63.2%$0.03152026-11-04source
24GPT-5.4 nano63%$0.006662026-11-04source
24Gemini 3.1 Flash-Lite63%$0.008172026-11-04source
24Gemini 2.5 Flash63%$0.01162026-11-04source
24Mistral Large 463%$0.01762026-11-04source
28Gemini 3.6 Flash60.7%$0.02262026-11-04source
29GPT-6 Luna60.1%$0.003022026-11-04source
29Claude Haiku 5.560.1%$0.003022026-11-04source
29Gemini 2.5 Flash-Lite60.1%$0.002772026-11-04source
29Mistral Small 460.1%$0.006752026-11-04source
29Codestral 25.0860.1%$0.01272026-11-04source
34Gemini 3.5 Flash-Lite57.7%$0.01162026-11-04source
35GPT-5 mini56.2%$0.009422026-11-04source
36GPT-4o mini45.2%$0.005312026-11-04source

LMArena Elo

33 of 55 verified models report this metric · unit: Elo · Higher is better · source: leaderboard

#ModelScorePer agent taskScore checkedSource
1Claude Opus 5.51,526$0.11682026-11-04source
2GPT-6 Astra1,525$0.30182026-11-04source
3Gemini 4 Argon1,523$0.05842026-11-04source
4Claude Sonnet 5.51,521$0.05842026-11-04source
5Claude Fable 5.11,520$0.28732026-11-04source
6GPT-6.1 Sol1,516$0.05842026-11-04source
7Claude Opus 51,512$0.15092026-11-04source
8GPT-6 Sol1,509$0.06042026-11-04source
9GPT-5.6 Sol1,508$0.12072026-11-04source
10Grok 4.71,507$0.05612026-11-04source
11GPT-5.6 Terra1,505$0.06542026-11-04source
11GPT-5.51,505$0.16342026-11-04source
11Gemini 3.8 Flash1,505$0.02262026-11-04source
14Gemini 3.1 Pro Preview1,504$0.06542026-11-04source
15DeepSeek Flash (V4.1)1,503$0.007842026-11-04source
16DeepSeek V4 Pro1,502$0.03152026-11-04source
17Claude Haiku 5.51,501$0.003022026-11-04source
17Gemini 3.6 Flash1,501$0.02262026-11-04source
19Claude Sonnet 51,485$0.06042026-11-04source
20GPT-5.41,465$0.08172026-11-04source
21GPT-5.4 mini1,461$0.02452026-11-04source
22Claude Sonnet 4.61,460$0.09052026-11-04source
23Gemini 2.5 Pro1,459$0.04712026-11-04source
24GPT-5.6 Luna1,451$0.006542026-11-04source
25Mistral Large 41,448$0.01762026-11-04source
26Gemini 3.1 Flash-Lite1,421$0.008172026-11-04source
27Gemini 2.5 Flash1,412$0.01162026-11-04source
28GPT-4.11,381$0.06112026-11-04source
29Claude Haiku 4.51,378$0.03022026-11-04source
30GPT-5 mini1,375$0.009422026-11-04source
31o4-mini1,362$0.03362026-11-04source
32Gemini 2.5 Flash-Lite1,337$0.002772026-11-04source
33GPT-4o mini1,289$0.005312026-11-04source

Humanity's Last Exam

24 of 55 verified models report this metric · unit: % · Higher is better · source: leaderboard

#ModelScorePer agent taskScore checkedSource
1Claude Opus 5.561.4%$0.11682026-11-04source
2Claude Fable 5.159.1%$0.28732026-11-04source
3Gemini 4 Argon57.1%$0.05842026-11-04source
4Claude Sonnet 5.555%$0.05842026-11-04source
5Claude Opus 554.9%$0.15092026-11-04source
6GPT-6 Astra54.7%$0.30182026-11-04source
7GPT-6.1 Sol52.9%$0.05842026-11-04source
8GPT-5.6 Sol49.5%$0.12072026-11-04source
9GPT-6 Sol47.9%$0.06042026-11-04source
10Gemini 3.8 Flash47.8%$0.02262026-11-04source
11Gemini 3.1 Pro Preview47%$0.06542026-11-04source
12GPT-6.1 Terra46.2%$0.02922026-11-04source
13GPT-5.545.8%$0.16342026-11-04source
14Claude Haiku 5.544.4%$0.003022026-11-04source
15Grok 4.743.1%$0.05612026-11-04source
16GPT-5.6 Terra42.9%$0.06542026-11-04source
17Gemini 3.6 Flash40.8%$0.02262026-11-04source
18GPT-6 Luna38.5%$0.003022026-11-04source
19Gemini 2.5 Pro22.5%$0.04712026-11-04source
20Gemini 3.5 Flash-Lite18.8%$0.01162026-11-04source
21Ministral 3 3B5.4%$0.003752026-11-04source
22Ministral 3 14B4.6%$0.00752026-11-04source
23Ministral 3 8B4.3%$0.005622026-11-04source
24Mistral Small 43.8%$0.006752026-11-04source

Output speed

24 of 55 verified models report this metric · unit: tok/s · Higher is better · source: leaderboard

#ModelScorePer agent taskScore checkedSource
1Gemini 3.8 Flash345$0.02262026-11-04source
2DeepSeek Flash (V4.1)217$0.007842026-11-04source
3GPT-5.6 Luna186$0.006542026-11-04source
4GPT-6 Sol153$0.06042026-11-04source
5Gemini 3.6 Flash148$0.02262026-11-04source
6Claude Sonnet 5.5139$0.05842026-11-04source
7Gemini 3.1 Pro Preview131$0.06542026-11-04source
8GPT-4.1120$0.06112026-11-04source
9GPT-5.6 Terra118$0.06542026-11-04source
10GPT-6.1 Terra109$0.02922026-11-04source
11Claude Sonnet 598$0.06042026-11-04source
12GPT-5.6 Sol96$0.12072026-11-04source
13Claude Opus 5.594$0.11682026-11-04source
14GPT-5.488$0.08172026-11-04source
15Gemini 2.5 Pro87$0.04712026-11-04source
16Grok 4.383$0.02982026-11-04source
17Gemini 4 Argon82$0.05842026-11-04source
18GPT-6.1 Sol74$0.05842026-11-04source
18Claude Opus 574$0.15092026-11-04source
20Claude Sonnet 4.673$0.09052026-11-04source
21GPT-5.570$0.16342026-11-04source
22Claude Fable 5.168$0.28732026-11-04source
23DeepSeek V4 Pro62$0.03152026-11-04source
24GPT-6 Astra53$0.30182026-11-04source

MMMU-Pro

20 of 55 verified models report this metric · unit: % · Higher is better · source: leaderboard

#ModelScorePer agent taskScore checkedSource
1Claude Opus 5.587.7%$0.11682026-11-04source
2GPT-6 Astra86.9%$0.30182026-11-04source
3GPT-6.1 Sol86%$0.05842026-11-04source
4Gemini 3.8 Flash85.6%$0.02262026-11-04source
5GPT-6.1 Terra84.7%$0.02922026-11-04source
5Claude Opus 584.7%$0.15092026-11-04source
7GPT-5.6 Sol83.4%$0.12072026-11-04source
8Gemini 3.6 Flash83.2%$0.02262026-11-04source
9GPT-6 Sol82.9%$0.06042026-11-04source
10Gemini 3.1 Pro Preview82.4%$0.06542026-11-04source
11GPT-5.6 Terra80.7%$0.06542026-11-04source
12GPT-5.579.9%$0.16342026-11-04source
13GPT-6 Luna79.7%$0.003022026-11-04source
14Gemini 3.5 Flash-Lite79%$0.01162026-11-04source
15Grok 4.778.2%$0.05612026-11-04source
16Gemini 2.5 Pro74.9%$0.04712026-11-04source
17Ministral 3 14B49.8%$0.00752026-11-04source
18Mistral Small 446.5%$0.006752026-11-04source
19Ministral 3 8B46%$0.005622026-11-04source
20Ministral 3 3B38.1%$0.003752026-11-04source

GPQA Diamond

15 of 55 verified models report this metric · unit: % · Higher is better · source: leaderboard

#ModelScorePer agent taskScore checkedSource
1GPT-6 Astra96.1%$0.30182026-11-04source
2Gemini 3.8 Flash95.3%$0.02262026-11-04source
3GPT-5.6 Sol94.1%$0.12072026-11-04source
3Gemini 3.1 Pro Preview94.1%$0.06542026-11-04source
5Claude Fable 5.193.7%$0.28732026-11-04source
6GPT-5.593.5%$0.16342026-11-04source
7Claude Opus 593.2%$0.15092026-11-04source
8Gemini 3.6 Flash92.8%$0.02262026-11-04source
9GPT-5.6 Terra92.5%$0.06542026-11-04source
10Gemini 2.5 Pro84.4%$0.04712026-11-04source
11Gemini 3.5 Flash-Lite83.8%$0.01162026-11-04source
12Ministral 3 14B57.2%$0.00752026-11-04source
13Mistral Small 457.1%$0.006752026-11-04source
14Ministral 3 8B47.1%$0.005622026-11-04source
15Ministral 3 3B35.8%$0.003752026-11-04source

Price vs performance

Reading the chart

Cheaper models sit to the left, higher-scoring models sit higher. Points in the lower-left are cheap and weak, points in the upper-right are expensive and strong, and the interesting ones sit above the trend: high score at a low task cost.

OpenAIAnthropicGooglexAIDeepSeekMistral
1,2701,3391,4081,4761,545$0.0100$0.1000Cost per typical agent task (USD, log scale)LMArena Elo (Elo)GPT-6 Astra — LMArena Elo 1,525; $0.3018 per typical taskGPT-6.1 Sol — LMArena Elo 1,516; $0.0584 per typical taskGPT-6 Sol — LMArena Elo 1,509; $0.0604 per typical taskGPT-5.6 Sol — LMArena Elo 1,508; $0.1207 per typical taskGPT-5.6 Terra — LMArena Elo 1,505; $0.0654 per typical taskGPT-5.6 Luna — LMArena Elo 1,451; $0.00654 per typical taskGPT-5.5 — LMArena Elo 1,505; $0.1634 per typical taskGPT-5.4 — LMArena Elo 1,465; $0.0817 per typical taskGPT-5.4 mini — LMArena Elo 1,461; $0.0245 per typical taskGPT-5 mini — LMArena Elo 1,375; $0.00942 per typical taskGPT-4.1 — LMArena Elo 1,381; $0.0611 per typical taskGPT-4o mini — LMArena Elo 1,289; $0.00531 per typical tasko4-mini — LMArena Elo 1,362; $0.0336 per typical taskClaude Fable 5.1 — LMArena Elo 1,520; $0.2873 per typical taskClaude Opus 5.5 — LMArena Elo 1,526; $0.1168 per typical taskClaude Opus 5 — LMArena Elo 1,512; $0.1509 per typical taskClaude Sonnet 5.5 — LMArena Elo 1,521; $0.0584 per typical taskClaude Sonnet 5 — LMArena Elo 1,485; $0.0604 per typical taskClaude Sonnet 4.6 — LMArena Elo 1,460; $0.0905 per typical taskClaude Haiku 5.5 — LMArena Elo 1,501; $0.00302 per typical taskClaude Haiku 4.5 — LMArena Elo 1,378; $0.0302 per typical taskGemini 4 Argon — LMArena Elo 1,523; $0.0584 per typical taskGemini 3.8 Flash — LMArena Elo 1,505; $0.0226 per typical taskGemini 3.6 Flash — LMArena Elo 1,501; $0.0226 per typical taskGemini 3.1 Pro Preview — LMArena Elo 1,504; $0.0654 per typical taskGemini 3.1 Flash-Lite — LMArena Elo 1,421; $0.00817 per typical taskGemini 2.5 Pro — LMArena Elo 1,459; $0.0471 per typical taskGemini 2.5 Flash — LMArena Elo 1,412; $0.0116 per typical taskGemini 2.5 Flash-Lite — LMArena Elo 1,337; $0.00277 per typical taskGrok 4.7 — LMArena Elo 1,507; $0.0561 per typical taskDeepSeek V4 Pro — LMArena Elo 1,502; $0.0315 per typical taskDeepSeek Flash (V4.1) — LMArena Elo 1,503; $0.00784 per typical taskMistral Large 4 — LMArena Elo 1,448; $0.0176 per typical taskGPT-4o miniGemini 2.5 Flash-LiteClaude Haiku 5.5
Cost = one typical agent task on the model at its standard list price; performance = LMArena Elo. Labels mark the highest score, the cheapest task and the lowest cost per elo point. Models without a published LMArena Elo score are not plotted. Sources are listed under Method below.

Cost per quality point

MetricModelsBest valueLowest $/pointMedian $/pointHighest $/point
AIME36Gemini 2.5 Flash-Lite$0.0000387$0.000380$0.003735
LiveCodeBench36Gemini 2.5 Flash-Lite$0.0000439$0.000456$0.004232
SWE-bench Verified36Gemini 2.5 Flash-Lite$0.0000460$0.000479$0.004451
Humanity's Last Exam24Claude Haiku 5.5$0.0000680$0.001305$0.005516
MMMU-Pro20GPT-6 Luna$0.0000379$0.000654$0.003472
GPQA Diamond15Ministral 3 3B$0.000105$0.000558$0.003140

Cost per point = the price of one typical agent task on the model divided by its score on that metric. It only works on scales that start at zero and reach 100, so it is not shown for Elo ratings or latency. Read it as "what one point of this score costs", not as a quality score of its own: a model can be the best value on one metric and mediocre on another.

What the ratio hides

A low cost per point can come from a cheap model with a weak score or from an expensive model near the top. Both can be the right buy depending on the bottleneck, so use the ratio to shortlist two or three models and then run your own prompts on them.

Method and sources

One table per metric

We never add two benchmarks into one combined score. A percentage, an Elo rating and a time are different scales, and averaging them would hide which one moved. Each table is sorted by its own metric only, and the column header says which direction is better.

Missing is missing

If a model has no published value for a metric it does not appear in that table. It is not shown as zero, it is not estimated from a sibling model, and it does not affect the ranks of the models that do report the metric.

Ties and ranks

Models with the same score share a rank; the next rank skips the tied places. Ranks are within the set of models that report that metric, not across the whole price list.

What "one agent task" means for the price column

Every price column prices the same task: 35K input tokens (55% cached) and 2,500 output tokens per task. That basis is median "typical task" token counts across the 14 tools in src/tools.json (35K input / 2,500 output, 55% cached). Rates are the standard list prices in the price table; batch and long-context tiers are not applied.

Where the scores come from

MetricUnitBetterModels reportingSourceData checked
AIME%higher36www.trustthebench.com2026-11-04
LiveCodeBench%higher36www.trustthebench.com2026-11-04
SWE-bench Verified%higher36www.trustthebench.com2026-11-04
LMArena EloElohigher33openlm.ai2026-11-04
Humanity's Last Exam%higher24artificialanalysis.ai2026-11-04
Output speedtok/shigher24www.swfte.com2026-11-04
MMMU-Pro%higher20artificialanalysis.ai2026-11-04
GPQA Diamond%higher15artificialanalysis.ai2026-11-04

Benchmark data compiled 2026-11-04; each model stores the actual read date in 'checked'. SWE-bench Verified, LiveCodeBench and AIME values are individual per-model rows from TrustTheBench (an aggregator that re-publishes standardized-protocol results, snapshot 2026-09-12, each row linking to the original benchmark source). TrustTheBench is tier-clustered: its 2026-09-12 snapshot reports the same rounded score for several models (e.g. 65.6 on SWE-bench Verified), so equal scores mean 'same reported tier', not identical capability, and small score differences are not meaningful. GPQA Diamond, MMMU-Pro and Humanity's Last Exam come from Artificial Analysis, which runs those evaluations itself; where a model publishes several reasoning-effort variants we use the highest-effort published variant and say so in the model 'notes'. LMArena Elo comes from the public Chatbot Arena+ leaderboard at openlm.ai (an aggregator snapshot, not the live lmarena.ai API); where that leaderboard lists a dated snapshot (e.g. GPT-4.1-2025-04-14) or a reasoning variant (e.g. GPT-5.5-high) for the same API model, we use it and record the exact entry in the model 'notes'. Output speed is the 'Speed' column of a public LMArena mirror (swfte.com/lmarena, aggregator), with a few values not present there taken from the BenchLeader coding leaderboard (per-model 'sources' override). Time-to-first-token and Aider Polyglot / BFCL are intentionally not published here: the accessible sources did not provide reliable per-model values (Artificial Analysis' latency figure for highest-effort reasoning variants is time-to-first-answer-token, not TTFT), and a missing metric is omitted rather than estimated, extrapolated or copied from a model family. Benchmarks are never comparable with each other - rank within one metric only.

Scores shared across models

Some metrics repeat the same rounded value across models: AIME: 4 repeated values covering 26 of 36 rows; LMArena Elo: 2 repeated values covering 5 of 33 rows; GPQA Diamond: 1 repeated value covering 2 of 15 rows; LiveCodeBench: 4 repeated values covering 26 of 36 rows; MMMU-Pro: 1 repeated value covering 2 of 20 rows; SWE-bench Verified: 4 repeated values covering 25 of 36 rows. An equal value means the source published the same reported figure, not that the models are equally capable on that metric, so we do not use these ties for fine-grained ranking between the models that share one.

For SWE-bench Verified, LiveCodeBench and AIME the values come from TrustTheBench, an aggregator that republishes results under a standardized protocol (mini-SWE-agent 2.0.0, snapshot 2026-09-12). Its published snapshot is tier-clustered: leading results are reported in the same band (its own page lists several models at 75.8%), so identical scores above indicate the same reporting band in that snapshot. We keep each model's own row from the source, but treat a shared value as a tie at the source rather than as evidence of equal capability.

Repeated LMArena Elo values come from the source's rounded ratings: the same Elo in the table is a tie in the published snapshot, not a separate measurement.

What this page does not claim

Leaderboards are not your product. A model that wins a public benchmark can still fail on your prompt, your data and your tool schema; treat the tables as a shortlist and run a small eval on real examples before you switch. Prices change, and the check date on each row is the only promise we make about a number.

Related tools, skills & plugins

Benchmarks tell you which model to shortlist; the token bill is decided by the agent, skill or plugin you run on it. These entries have a cost page or an official site, and the hub lists every entry by category.

Related

Questions

Which LLM is best overall?

There is no overall winner because the benchmarks measure different skills. The highest LMArena Elo score is Claude Opus 5.5 at 1,526, while the cheapest model per typical agent task is Gemini 2.5 Flash-Lite at $0.00277. Pick the metric that matches your workload and then compare cost at your own token counts.

Why not publish one combined benchmark score?

Because a combined score needs weights, and any weighting hides the trade-off you are buying. Keeping one table per metric lets you apply your own weights: a coding agent and a document summarizer do not need the same ranking.

Where do the benchmark numbers and dates come from?

Each metric names its source in the source table at the bottom of this page and every row shows the date that score was checked. Primary leaderboards are used first; when a figure comes from an aggregator that is stated in the note.

Why is a model missing from one of the tables?

That model does not report the metric, so it is left out rather than shown as zero or estimated. The table header shows how many of the verified models report each metric.

How do you compute cost per quality point?

Divide the list-price cost of one typical agent task on the model by its score on a 0-100 metric. It is only shown for metrics with a zero-to-100 scale; Elo ratings and latency are excluded because a per-point ratio would not mean anything.

Are these scores independent measurements of every model?

Not always. Some sources repeat the same rounded score across models: the SWE-bench, LiveCodeBench and AIME values here come from a standardized-protocol snapshot that is tier-clustered, so several models can sit in the same reported band. The shared-scores note under Method lists every metric and row count affected; treat a repeated value as a tie at the source, not as evidence of equal capability or a ranking gap.

How current is this page?

The benchmark data was last read on 2026-11-04 and each row keeps its own check date; prices were last checked on 2026-11-18. When a source changes a score we update the row and move its date, and a score without a readable source is removed rather than kept as a stale number.

Next step → See what one task costs in tokens

Price data last checked 2026-11-18 against official sources · how we verify

Price-change alerts

One short email when a provider changes a price or a promo ends. No spam, unsubscribe any time.