Skip to content

Cost per task vs price per token

A model's list price is per token, but your bill depends on how many tokens it spends on each job. Wordy answers and long hidden reasoning can make a cheap-looking model expensive, and a terse model cheaper than its price suggests. Cost per task is the average spend on one task of the Artificial Analysis Intelligence Index, counting the tokens each model actually used, as of Nov 16, 2026. 52 models have both figures.

Biggest surprises

Pricier than they look

Rank much worse per task than by list price: they spend more tokens per job.

Cheaper than they look

Rank much better per task than by list price: they finish jobs in fewer tokens.

All models, cheapest per task first

Model$/taskBlended $/M tokensRank by price → per taskAnalyzer Index
Mistral Small 4
Mistral
$0.04$0.26#12 → #1—
GPT-6 Luna
OpenAI
$0.04$0.10#1 → #263.7
Claude Haiku 5.5
Anthropic
$0.09$0.20#4 → #361.9
Muse Ember
Meta
$0.12$0.20#5 → #4—
MiMo-V2.6-Flash
Xiaomi (MiMo)
$0.13$0.18#3 → #5—
GPT-6.1 Terra
OpenAI
$0.17$1.00#23 → #6—
Muse Glimmer 30B
Meta
$0.18$0.52#17 → #7—
Gemini 3.1 Flash Lite
Google
$0.19$0.28#13 → #850.1
Inkling Small
Thinking Machines
$0.20$0.64#20 → #959.4
MiMo-V2.6-Pro
Xiaomi (MiMo)
$0.26$0.54#18 → #10—
Nemotron 3.5 Lightning
NVIDIA
$0.31$0.10#2 → #11—
GPT-5.6 Luna
OpenAI
$0.33$0.23#6 → #1267.5
gpt-oss-120b
OpenAI
$0.35$0.26#10 → #1344.3
GPT-6.1 Sol
OpenAI
$0.36$2.00#34 → #1488.7
DeepSeek V4.1 Flash 1017
DeepSeek
$0.46$0.26#9 → #15—
MiniMax M3.1 Flash
MiniMax
$0.49$0.26#11 → #16—
Claude Haiku 4.5
Anthropic
$0.53$2.00#30 → #1745.5
Nemotron 3.5 Super VL
NVIDIA
$0.56$0.24#8 → #18—
GLM 5.3 Flash
Zhipu AI (GLM)
$0.58$0.24#7 → #1969.8
Gemini 3.5 Flash Lite
Google
$0.69$0.43#14 → #2044.4
Claude Sonnet 5.5
Anthropic
$0.78$4.00#44 → #2181.5
Gemini 4 Hydrogen
Google
$0.80$1.50#28 → #22—
Hy4
Tencent (Hunyuan)
$1.03$0.63#19 → #23—
Mistral Medium 3.5
Mistral
$1.04$3.00#42 → #24—
Inkling 2 Small
Thinking Machines
$1.18$0.48#15 → #25—
MiniMax M3
MiniMax
$1.29$0.52#16 → #2653.5
Inkling
Thinking Machines
$1.35$1.76#29 → #2759.2
Mistral Large 4
Mistral
$1.41$1.05#24 → #2861.9
GLM 5.3
Zhipu AI (GLM)
$1.57$2.15#37 → #2972.4
DeepSeek V4 Pro 0423
DeepSeek
$1.77$1.05#26 → #3064.3
Muse Spark Open
Meta
$2.02$2.00#36 → #31—
Qwen3.8 27B
Alibaba (Qwen)
$2.16$0.96#22 → #32—
Nemotron 3 Ultra
NVIDIA
$2.21$1.05#25 → #3357.9
Step 5 Preview
StepFun
$2.35$1.43#27 → #34—
Kimi K3
Moonshot (Kimi)
$2.37$6.00#46 → #3577.3
Muse Spark 1.3
Meta
$2.39$2.00#35 → #3687.0
Claude Opus 5.5
Anthropic
$2.45$8.00#47 → #3793.8
GPT-6 Sol
OpenAI
$2.47$2.00#33 → #3879.8
GPT-6 Astra
OpenAI
$2.73$10.00#50 → #3991.3
Qwen4 Max
Alibaba (Qwen)
$3.27$2.25#39 → #40—
Kimi K3.1 Pro
Moonshot (Kimi)
$3.68$4.00#45 → #4181.5
GPT-5.6 Terra
OpenAI
$3.73$2.25#38 → #4277.5
Claude Fable 5.2
Anthropic
$3.80$16.00#51 → #43—
Gemini 4 Argon
Google
$4.31$2.00#31 → #4492.8
Muse Wonder
Meta
$4.56$8.00#48 → #4589.0
Gemini 4 Carbon
Google
$4.68$2.00#32 → #4695.7
Grok 4.6
xAI
$4.68$3.00#40 → #4774.3
Claude Fable 5.1
Anthropic
$5.56$20.00#52 → #4891.3
Grok 4.7
xAI
$6.18$3.00#41 → #4971.3
Gemini 3.8 Flash
Google
$6.26$0.75#21 → #5081.0
Qwen3.8 Max (0902)
Alibaba (Qwen)
$10.02$3.00#43 → #51—
Claude Opus 5
Anthropic
$11.82$10.00#49 → #5286.9

Reading these figures

Blended price assumes three input tokens per output token. Cost per task reflects one task mix (Artificial Analysis's evaluation suite); your prompts may use more or fewer tokens. Effort variants of the same model are listed separately because higher reasoning effort uses more tokens per task. For your own workload, use the cost calculator, or sort the leaderboard by $/task.