BenchLeader

Which model should you use?

Say how much quality, speed and cost matter, what the model has to be good at, and what you cannot do without. The answer updates as you go: a top pick, the runners-up, why each is where it is, and what a cheaper or faster choice would cost you in quality. Every score comes from independent benchmarks; nothing a model's maker published about itself is used.

331 ranked models · list prices and speeds · data as of 16 Nov 2026

Quality first, cost second, speed a tiebreak; about 36M tokens a month.

What matters
High
Low
Some
Good at

Quality is the mean of what you pick, from independent results only.

Your volume, per day
A number, optionally followed by k, M or B.
A number, optionally followed by k, M or B.

Count thinking in tokens out: reasoning models bill it as output.

Must have
$
Reasoning

Images and audio: models an independent publisher has tested on that input.

Providers: All providers

Recommendation

250 of 331 ranked models fit. Ranked on the Index (57%), speed (14%) and cost (29%) at 36M tokens a month.

Top pick

Gemini 4 CarbonhighGoogleSolid evidence. Index ±1.9 (approximate 95% band).

#2 on the Index of the 250 models that fit: the one above it costs at least 4× as much.

Index
68.9±1.9
A month at your volume
$106
Per 1M at your mix
$2.94
Output speed
64 tok/s
First answer
73 s
Context
1.0M

Model pageCompare the top 3Cost calculator

Runners-up

  1. 2Muse Spark 1.1MetaSome evidence. Index ±5.6 (approximate 95% band).

    #23 on the Index, 6.7 points behind the top pick, at half the top pick's price, 3× the output speed.

    Index
    62.2±5.6
    A month at your volume
    $55
    Output speed
    193 tok/s
    First answer
    1.8 s
    Context
    1.0M
  2. 3Gemini 3.7 FlashhighGoogleSolid evidence. Index ±3.5 (approximate 95% band).

    #24 on the Index, 6.8 points behind the top pick, 2.6× cheaper than the top pick, 1.5× the output speed.

    Index
    62.1±3.5
    A month at your volume
    $40
    Output speed
    95 tok/s
    First answer
    2.2 s
    Context
    1.0M
  3. 4Claude Fable 5.2maxAnthropicSome evidence. Index ±4.3 (approximate 95% band).

    #1 on the Index, 3.8 points ahead of the top pick, 1.7× the output speed, 4× the price of the top pick.

    Index
    72.7±4.3
    A month at your volume
    $421
    Output speed
    109 tok/s
    First answer
    4 min
    Context
    1M
  4. 5Muse Spark 1.3MetaThin evidence. Index ±8.2 (approximate 95% band). The index rests on few independent results so far.

    #27 on the Index, 7.3 points behind the top pick, at half the top pick's price, 2.5× the output speed.

    Index
    61.6±8.2
    A month at your volume
    $55
    Output speed
    163 tok/s
    First answer
    2.6 s
    Context
    1.0M
The top 40 of the 250 that fit
#ModelIndexA monthSpeedFirst answerContext
1Gemini 4 Carbonhigh68.9$1066473 s1.0M
2Muse Spark 1.162.2$551931.8 s1.0M
3Gemini 3.7 Flashhigh62.1$40952.2 s1.0M
4Claude Fable 5.2max72.7$4211094 min1M
5Muse Spark 1.3thin61.6$551632.6 s1.0M
6Gemini 4 Hydrogenhigh63.4$7917016 s1.0M
7Gemini 3.6 Flashhigh61.0$4017711 s1.0M
8DeepSeek V4 Flashhigh57.2$7.0820.7 s1.0M
9Muse Spark 1.2xhigh61.5$5520811 s1.0M
10Qwen3.6 Maxmax62.1$77––262k
11Gemini 3 Flashthinkingthin58.9$301964.7 s1.0M
12Gemini 3.5 Flashhigh62.5$891769.9 s1.0M
13GPT-6.1 Terramax64.1$53723 min1.1M
14Gemma 4 26B A4B56.1$7.1760.4 s262k
15GPT-5 nanothin56.2$3.61062.9 s400k
16Claude Opus 5.5max68.4$2121166 min1M
17Claude Haiku 5.5xhigh56.7$5.323150 s1M
18Gemini 4 Argonhigh65.3$1066646 s1.0M
19GPT-6.1 Solmax66.6$106503 min1.1M
20GPT-5.6 Lunaxhigh58.2$1211028 s1.1M
21Grok 4.20 Multi-Agentthin58.5$4534912 s2M
22Gemini 3.8 Flashmedium60.2$4011615 s1.0M
23DeepSeek V4.1 Flash 1017max57.7$1417440 s1.0M
24Beam 1.0 501B A23B54.2$131930.3 s1.0M
25Step 3.5 Flash56.3$4.213325 s256k
26Claude Sonnet 5.5max64.2$1061525 min1M
27Qwen3.5 Plusthinking57.4$26800.8 s1M
28Grok 4.1 Fastthinking55.3$7.9––128k
29Inkling 2 Small57.9$1310049 s1.0M
30GPT-6 Solmax62.6$10612737 s1.1M
31Muse Wonderthin64.4$212703.3 s1.0M
32GPT-5.6 Solmax65.3$21310034 s1.1M
33MiMo-V2.556.3$3.8282.5 s1M
34GLM 5.3 Flash57.8$6.65487 s1M
35GLM-5thinking60.2$436929 s205k
36Muse Embermax52.1$5.34375.0 s1.0M
37o360.4$971315.2 s200k
38GPT-5 mini56.6$18703.5 s400k
39MiMo-V2-Flashthinking53.9$4.8––256k
40Nemotron 3 Ultra 550B A55B55.4$321060.3 s1M

How it ranks

Requirements come first and are absolute: a model over your budget, slower to its first answer than you allow, or without a result on image input when you need one, is out. What is left is ranked on three things. Quality is the mean of the capabilities you picked (category scores of the BenchLeader Index, where 50 is the average evaluated model and 15 points is one standard deviation). Speed is output tokens per second and the wait for the first answer. Cost is what a month at your volume comes to on list prices, input, cached input and output each at its own rate.

Each is measured in standard deviations across the models that fit, with speed and cost on a log scale, and the sliders weigh them: each step up doubles the weight. A bill under about $20 a month counts as cheap whatever the model, so pennies never outvote quality. Each model is represented by its best reasoning-effort setting for your weights, and a model needs most of the weighting measured, and a price when cost counts, to be ranked at all.

The same weighting, fixed and printed, is behind the best model by use case pages. To filter without weighting, use find a model; to price one model across plans and usage patterns, the cost calculator.

Questions people ask

How do I know how many tokens I use a day?
Your API dashboard (OpenAI, Anthropic, Google, OpenRouter) shows input and output tokens per day. As a guide, a support chatbot with a few thousand conversations a day uses one to five million input tokens, and a coding agent running most of a workday reads ten to twenty million, nearly all of it from the prompt cache.
What does thin evidence mean?
The model's score rests on few independent results so far, or on results that disagree with each other a lot, so its position could move as more boards test it. The band beside the Index (±) is an approximate 95% interval.
Why is a reasoning-effort setting recommended rather than a model?
The same model at a higher effort setting usually scores higher and answers more slowly, and bills its thinking as output tokens. The recommender ranks every setting it has results for and shows the one that suits your weights best; the label beside the name says which.
Can I share a recommendation?
Yes. The settings are kept in the page address after the #, so copying the link copies the recommendation, and it re-ranks on the day's data whenever it is opened.