Skip to content
InfratoApp
InfratoApp / LLM price index

LLM price index

A single lab’s list prices span up to 8,333× — OpenAI runs $0.018 to $150.00 per 1M input tokens. Medians, not means, per lab below.
⚑ What this is not, yet: a settled index. These are the list prices of 16 sellers, grouped by the lab that made each model. Group medians move when a seller starts or stops listing a model, not only when a price changes.
Zero-price (free) models are excluded from these ranges — 23 left out · see every model we price →

The market at a glance

700 models · 28 labs

Token prices fall as capability moves down the stack, so an index has to be measured per model: today’s rate against the same model’s rate then. A model counts only if it was priced on both dates, which keeps launches and retirements out of the figure. Each seller is counted once per model.

$0.56
Median input rate · 700 models
28
Labs with a priced model
○
Past 4 weeks
needs the same model priced on two dates
○
Past year
needs a year of readings
Closed weights vs open weightsBLENDED $/1M
○ Splitting these needs a weights licence for each model, which none of the price feeds we read carries. Guessing it from the maker’s name would be wrong for exactly the models people argue about.
What a cheaper tier buysVS THE SAME SELLER’S FULL INPUT RATE
Only 12 of 16 sellers publish a cached input rate at all — the rest charge the full input rate on a cache hit, which is the difference between a prompt cache saving money and saving nothing.

Repricing history — blended $/1M over time

WAITING

A repricing line needs the same model priced on two dates. Token prices are re-read, but the series is not yet long enough for a move to mean anything — and a chart drawn through one reading per model would assert a direction that nothing in the data supports. This is the block that will be most read once it works, which is exactly why it must not be faked first.

NEEDS THE SAME MODEL PRICED TWICEHAVE 700 MODELS · ONE SERIES

Per-lab ranges today

Blended $/1M tokens · medians, not means · open a row for the models behind the range

LABIN LOWIN MEDIN HIGHMODELS
InclusionAIout $0.063–$0.225, typically $0.200$0.021$0.068$0.075
IInceptionout $0.150–$0.750, typically $0.350$0.040$0.100$0.250
Tencentout $0.177–$2.50, typically $0.528$0.044$0.132$0.834
UUpstageout $0.200–$0.600, typically $0.600$0.050$0.150$0.150
Metaout $0.050–$20.00, typically $0.590$0.020$0.188$4.00
NVIDIAout $0.200–$2.40, typically $0.500$0.050$0.200$0.600
Stepfunout $0.300–$2.70, typically $1.15$0.100$0.200$1.00
Alibabaout $0.130–$12.00, typically $1.00$0.030$0.220$4.00
Mistralout $0.030–$16.00, typically $0.750$0.029$0.250$16.00
ByteDanceout $0.200–$3.00, typically $2.00$0.075$0.250$0.500
Deepseekout $0.030–$8.00, typically $1.00$0.020$0.257$3.00
Amazonout $0.140–$12.50, typically $2.50$0.035$0.300$2.50
Cohereout $0.150–$10.00, typically $1.50$0.037$0.300$2.50
Minimaxout $0.600–$2.20, typically $1.20$0.150$0.300$0.550
SSao10kout $0.050–$0.850, typically $0.400$0.040$0.350$0.850
Googleout $0.100–$20.00, typically $2.50$0.050$0.375$2.00
TThedrummerout $0.400–$0.800, typically $0.500$0.300$0.400$0.550
Xiaomiout $0.280–$8.70, typically $0.870$0.140$0.435$4.35
Z.aiout $0.080–$8.80, typically $1.98$0.029$0.540$2.80
Moonshotout $2.30–$11.50, typically $3.41$0.179$0.600$2.28
AAion-labsout $1.40–$6.00, typically $1.60$0.700$0.800$3.00
UUnbiasedout $3.20–$7.50, typically $3.20$0.800$0.800$2.50
Nous Researchout $0.700–$3.00, typically $1.00$0.700$1.00$1.00
OpenAIout $0.011–$600.00, typically $5.00$0.018$1.10$150.00
xAIout $2.00–$6.00, typically $2.50$1.00$1.25$2.00
Perplexityout $1.00–$15.00, typically $8.00$1.00$2.00$3.00
Anthropicout $0.250–$75.00, typically $15.00$0.050$3.00$15.00
Sakanaout $4.00–$30.00, typically $18.00$0.950$3.50$5.00

Frequently asked

What does a million tokens cost?
It depends far more on which model than on which lab. Inside a single maker’s own catalogue the cheapest and dearest model can differ by thousands of times per million input tokens, because a small model and a frontier model are different products sold under one brand.
Why are input and output priced differently?
Output costs the seller more to produce: every output token is generated one at a time, while input is processed in parallel. Output typically lists at three to five times input, which is why we never average the two into one number.
Is this a cross-vendor index?
Not while the prices come from one seller. What is published here is each lab’s own range — cheapest model to dearest — which is a fact about a product line rather than about a market. Calling it an index before a second independent seller exists would be the wrong word for the data.
Why are free models excluded from the ranges?
A model given away at zero is not priced, it is promoted. Including it would drag every range down to $0.00 and make a lab that gives one model away look categorically cheaper than one that does not.
What is a cached input rate?
A reduced price for input tokens the seller has already processed in a recent request, usually a long system prompt reused across calls. Not every seller publishes one, and the ones that do not charge the full input rate on a cache hit — so it is listed per model rather than assumed.