Skip to content
InfratoApp
Tools / LLM cost calculator

What your workload costs, per model

700 models carry both an input and an output token price, read from each provider's own page. The lowest combined rate on record is scaledown/extract at $0.04 per million input tokens and $0 per million output tokens.

LLM $/1M tokGPU $/hr

700 token prices across 272 models. Each row is a seller’s listing, so the same model from four sellers is four rows. Input and output are priced separately, and a model appears only when both are on record. 31 carry one half and are left out.

Model
Serving tier
Data residency · NOT PUBLISHED BY THESE SELLERS
USEUAPAC
Provisioned tier · SOLD BY THE GPU-HOUR, NOT THE TOKEN
Provisionedsee GPU prices →

15 of 15 listings match · 2 models selected · no cached tier → billed at full input rate · batch is half the standard rate where we hold both

Cheapest for this workload: z-ai/glm-5.3 from OpenRouter — $64.39 a month for your mix.

Model · seller$/mo ↓vs cheapestRead from
z-ai/glm-5.3OpenRouter · standard$64.39—2026-10-27
z-ai/glm-5.3OpenRouter · batch$134.00+108%2026-10-27
z-ai/glm-5.3Baseten · standard$344.00+434%2026-10-22
z-ai/glm-5.3Fireworks AI · standard$344.00+434%2026-10-22
z-ai/glm-5.3FriendliAI · standard$344.00+434%2026-10-21
z-ai/glm-5.3Novita AI · standard$344.00+434%2026-10-22
z-ai/glm-5.3SambaNova · standard$344.00+434%2026-10-22
z-ai/glm-5.3Together AI · standard$344.00+434%2026-10-27
moonshotai/kimi-k3OpenRouter · standard$482.68+650%2026-10-27
moonshotai/kimi-k3OpenRouter · batch$729.60+1033%2026-10-27
moonshotai/kimi-k3Baseten · standard$960.00+1391%2026-10-22
moonshotai/kimi-k3Fireworks AI · standard$960.00+1391%2026-10-22
moonshotai/kimi-k3Moonshot AI · standard$960.00+1391%2026-10-22
moonshotai/kimi-k3Novita AI · standard$960.00+1391%2026-10-22
moonshotai/kimi-k3Together AI · standard$960.00+1391%2026-10-27

Tokens × the rate each seller published · nothing here is averaged. Every seller resells somebody else’s model, so there is no first-party filter to offer · what we collect

Frequently asked

What's the cheapest way to run a model per million tokens?
scaledown/extract at OpenRouter: $0.04 in and $0 out per million, the lowest input rate among 700 models we hold both halves for. Cheapest is not the same as suitable — a model that needs three attempts at your task costs three times its rate.
How is blended cost computed?
Fresh input × the input rate, plus cached input × the cached rate where the seller publishes one and the full input rate where they do not, plus output × the output rate. Nothing is weighted and nothing is averaged: input and output are different products at different prices, and a single blended rate hides which one your workload is actually buying.
What does a cached tier change?
471 of the models here publish a cached input rate. It applies to a prefix the seller has already processed — usually a long system prompt reused across calls. Where a seller publishes none, a repeated prefix bills at the full input rate, which is the difference between a prompt cache saving money and saving nothing.
How many tokens is my workload?
Roughly four characters to a token, so a million tokens is about 750,000 words. A back and forth of a few paragraphs is around 5,000 tokens including the system prompt, which puts a thousand conversations a day near 150 million a month. Output is usually the smaller half by volume and the dearer half by rate.
Is the cheapest model the cheapest answer?
Not necessarily, and this page cannot tell you. A model that needs three attempts at your task costs three times its rate, and we hold no benchmark that would let us say which models those are. The rates here are exact; whether a model is suitable is a judgement this directory does not make.
Can a seller pay to rank cheapest here?
No. The ranking is the arithmetic of the rates we read against the workload you entered, and nothing in it can be bought. How the site is paid for is set out under money and independence at /our-data#sec5.

What this figure does not include

PRICES ARE LIST RATES READ FROM EACH PROVIDER AND ARE AT MOST 21 DAYS OLD. CACHED INPUT IS MODELLED AT THE SHARE YOU SET, WHERE THE SELLER PUBLISHES A RATE. NEGOTIATED DISCOUNTS ARE NOT, SO A REAL BILL IS USUALLY LOWER THAN THIS. BATCH RATES ARE INCLUDED WHERE A SELLER PUBLISHES ONE, AS A TIER YOU CAN SELECT.