The dearest AI model costs 6,000 times more than the cheapest
A million tokens of output costs three cents from the cheapest model we track and $180 from the dearest. Both are for sale today. The gap is real, but the two ends are not selling the same thing.
Mistral Nemo will write you a million tokens for $0.03. GPT-5.5 Pro charges $180 for the same amount. On OpenAI's rule of thumb that 100 tokens is about 75 words, a million tokens is roughly 750,000 words. So the same pile of text costs three cents or $180 depending on which model writes it.
That is a ratio of 6,000 to one, across the 65 language models we list as active and priced on 14 November 2026.
The ladder
Sort the 65 models by output price and take every fifth one, and you get the ladder below. The scale is logarithmic, which means each gridline is ten times the one before. On an ordinary scale everything under $10 would be squashed flat against the left edge.
Output price per million tokens, every fifth model from cheapest to dearest
The median output price is $7.50 per million tokens and the median input price is $1.25.
Where the models sit
Active language models by output price band
The spread is fairly even until the top. 17 of the 65 models charge under a dollar per million output tokens. Only 6 charge $50 or more: five "Pro" tiers from OpenAI and Anthropic's Claude Fable 5.
| Five cheapest | Input | Output | Price from |
|---|---|---|---|
| Mistral Nemo (Mistral) | $0.03 | $0.03 | Cheapest host on OpenRouter |
| Llama 3.1 8B Instruct (Meta) | $0.02 | $0.04 | Cheapest host on OpenRouter |
| Phi 4 (Microsoft) | $0.07 | $0.14 | Cheapest host on OpenRouter |
| Nova Micro 1.0 (Amazon) | $0.04 | $0.14 | Cheapest host on OpenRouter |
| Command R7B (12-2024) (Cohere) | $0.04 | $0.15 | Cheapest host on OpenRouter |
| Five dearest | Input | Output | Price from |
| GPT-5.5 Pro (OpenAI) | $30 | $180 | Lab's own list |
| GPT-5.4 Pro (OpenAI) | $30 | $180 | Lab's own list |
| GPT-5.2 Pro (OpenAI) | $21 | $168 | OpenRouter listing |
| GPT-5 Pro (OpenAI) | $15 | $120 | OpenRouter listing |
| o3 Pro (OpenAI) | $20 | $80 | OpenRouter listing |
The two ends are not like for like
Look at the "price from" column. The cheap end is made up of small models such as Mistral Nemo, Llama 3.1 8B and Phi 4 that anyone may host, priced at whichever host on OpenRouter is cheapest that day. The dear end is a lab's own list price for its top tier. Comparing them is a bit like comparing the cheapest supermarket's own-brand with a restaurant's tasting menu. Both are food and the price gap is true, but nobody thinks they are the same product.
A fairer test is to use only prices taken from the labs' own pricing pages. We hold 20 of those among the active models. They run from GPT-5.4 Nano at $1.25 to GPT-5.5 Pro at $180, which is 144 to one. Both of those are OpenAI models, so one lab's own price list spans more than a hundredfold.
| Lab | Models | Cheapest | Dearest | Range |
|---|---|---|---|---|
| OpenAI | 21 | $0.40 | $180 | 450x |
| Anthropic | 10 | $5 | $50 | 10x |
| Meta | 6 | $0.04 | $20 | 500x |
| Cohere | 4 | $0.15 | $10 | 67x |
| 10 | $0.40 | $10 | 25x | |
| Mistral | 4 | $0.03 | $6 | 200x |
| Amazon | 3 | $0.14 | $3.20 | 23x |
Anthropic stands out in that table for a different reason. Its cheapest model in this set, Claude Haiku 4.5, costs $5 per million output tokens, which is more than the cheapest from every other lab listed.
What the gap does not prove
It would be neat to finish by saying whether the $180 model is 6,000 times better. We cannot say, and we would be wary of anyone who claims to. Price per token is easy to measure. Quality is not, and we do not yet hold independent test results we are allowed to publish alongside these prices. Our transparency page explains what is missing and why.
There is a second gap in this comparison. Price per token is not cost per job. A model that reasons at length before it answers uses more output tokens to do the same task. OpenAI's documentation (read on 14 November 2026) says reasoning tokens "are billed as output tokens" and can run to tens of thousands per response. So a higher price can be multiplied by a higher token count. Two models with the same list price can produce very different bills.
Start at the cheap end
If the job is simple and the volume is high, such as sorting messages into categories, pulling fields out of documents or writing short summaries, try a model from the bottom third of the ladder first and move up only if the results are not good enough. Trying the cheap model on a small sample first costs very little, and you will know from the output whether you need to pay more.
For the sums on your own workload, use the cost calculator. The guide to AI pricing covers input, output and cached tokens, and the cheapest models page keeps a current list.