What 108 AI models cost per million tokens

One dot per model, by provider, on the same log scale; each gridline is ten times the last. The short vertical tick is each provider's median. OpenAI has the widest range, $0.50 to $600, a 1,200× spread inside one price list.

$10 of output from ministral-3-3b

100M

$10 of output at the median price

2.1M

$10 of output from o1-pro

16,667

Current price per million output tokens. The top ten: OpenAI 7, Anthropic 3. None of the ten cheapest comes from OpenAI, Anthropic, Google and xAI.

Divide each model's output price by its input price and the median is 5×; 34 of 108 models sit exactly there. 19 of those 34 are Anthropic: every one of its 19 models charges exactly 5× input. xAI prices all 8 of its models at 2× or 3×. The six models that charge the same for output as input are all Ministral and Llama models. The three Llama prices come from Together AI, not from Meta, and the dataset flags them for review.

The headline price is rarely the whole bill. These are the rules the other price types follow, each one checked against every model that lists it.

Of the 83 models that list a cache-read price, 60 charge 1/10 of input, 12 of them Alibaba rates derived from its published 10% rule. Every cache price on a model first recorded before May 2025 is OpenAI's, and none is below 1/4. Of the models recorded since, 9 charge more than 1/10, and the seven that charge less were all first recorded from June 2026 on. These are today's prices: the index holds no cache-read price older than 31 March 2026, so the chart shows which models charge what now, not how cache prices moved.

The index records 26 changed price rows, but they belong to ten repricings, each one model on one date. OpenAI's change to gpt-5.6-sol on 22 August 2026 moved eight rates at once. Seven repricings were cuts and three were rises, all three in 2026 from Mistral and DeepSeek. The deepest cut is o3, from $40 to $8 per million output tokens. Four of the ten fall between 13 and 22 August 2026, the only month with more than one. The index first recorded 70 of its 130 models in 2026, so an older repricing is less likely to be on record. Three of the changes need context, and the note under each chart gives it.

22 of the 130 models have left their price lists, 19 of them in 2026. The median one was listed for 444 days; gpt-4 lasted 1,263. Six of them end on the same day, 28 August 2026, because that is when the collector noticed they were gone, not when they went. Leaving the list is not the same as being switched off: gpt-4 and gpt-4-turbo are gone from OpenAI's price list, but the dataset's notes give both a scheduled retirement date of 23 October 2026.

Three public datasets are pulled here. The AI Model Price Stability dataset reads prices from web-archive captures of the Anthropic, Google and OpenAI pricing pages, with a link to every capture. LLM Token Price History keeps a current table for 35 providers, resellers included. Where either states a price the index also holds, for the same model and price type, they match 269 times out of 276: 60 of 60 archive prices on the day of the capture, and 209 of 216 current prices across 83 models. Of the 83 shared models, 13 carry a non-current status in LLM Token Price History (8 deprecated, 4 retired, 1 preview) while the index still lists them; the prices agree anyway.

Price per million tokens. Set the size of a call to see what it costs on each model. Which model is cheap depends on the shape of the call. From a long document (50,000 tokens in, 1,000 out) to a long answer (1,000 in, 8,000 out), 17 of 108 models move ten or more places in the order. llama-3.1-405b moves 35 places towards the cheapest, because it charges the same for output as for input. Its prices are Together AI's, flagged for review in the dataset.