ThinkFacility

What the AI models actually cost

List prices for 48 models from 8 labs, on one page. Put your own numbers in the boxes and every row re-prices itself, so you can see what your week would cost on each one.

Prices per million tokens, read off each lab's own docs · updated 2026-09-05 12:15 UTC

What would this cost me?

Say what one request looks like and how often you run it. We count words, then turn them into tokens at the usual English rate.

Cheapest for that workload: Ministral 3 14B at $0.19 a month.

Model Lab Input Output Context Your month
Leanstral 1.5MistralMistralFreeFree256K$0
Ministral 3 14BMistralMistral$0.20$0.20256K$0.19
Qwen3.8-FlashAlibabaAlibaba$0.15$0.471M$0.27
Mistral Small 4MistralMistral$0.15$0.60256K$0.32
CodestralMistralMistral$0.30$0.90128K$0.53
GPT-5.6 LunaOpenAIOpenAI$0.20$1.201M$0.59
GPT-5.4 nanoOpenAIOpenAI$0.20$1.25400K$0.61
Gemini 3.1 Flash-LiteGoogleGoogle$0.25$1.501M$0.74
DeepSeek-V4-FlashDeepSeekDeepSeek$0.44$1.321M$0.77
DeepSeek-V4-Flash-Vision-ExpDeepSeekDeepSeek$0.44$1.321M$0.77
Qwen3.7-PlusAlibabaAlibaba$0.40$1.601M$0.86
Mistral Large 3MistralMistral$0.50$1.50256K$0.88
Gemini 3.5 Flash-LiteGoogleGoogle$0.30$2.501M$1.17
Grok Build 0.1xAIxAI$1.00$2.00256K$1.36
Gemini 3 Flash PreviewGoogleGoogle$0.50$3.001M$1.48
Grok 4.3xAIxAI$1.25$2.501M$1.70
Gemini 3.6 FlashGoogleGoogle$0.75$3.751M$1.92
Gemini 3.7 FlashGoogleGoogle$0.75$3.751M$1.92
Gemini 3.8 FlashGoogleGoogle$0.75$3.751M$1.92
GPT-5.4 MiniOpenAIOpenAI$0.75$4.50400K$2.22
DeepSeek-V4-ProDeepSeekDeepSeek$1.32$3.961M$2.32
Muse Spark 1.1MetaMeta$1.25$4.251M$2.40
Muse Spark 1.2MetaMeta$1.25$4.251M$2.40
Muse Spark 1.3MetaMeta$1.25$4.251M$2.40
Claude Haiku 4.5AnthropicAnthropic$1.00$5.00200K$2.56
Grok 4.5xAIxAI$2.00$6.00500K$3.52
Grok 4.6xAIxAI$2.00$6.00500K$3.52
Qwen3.8-MaxAlibabaAlibaba$2.00$6.001M$3.52
Mistral Medium 3.5MistralMistral$1.50$7.50256K$3.84
Gemini 3.5 FlashGoogleGoogle$1.50$9.001M$4.44
Gemini 2.5 ProGoogleGoogle$1.25$10.001M$4.70
GPT-5.1OpenAIOpenAI$1.25$10.00400K$4.70
Claude Sonnet 5AnthropicAnthropic$2.00$10.001M$5.12
Gemini 3.1 Pro PreviewGoogleGoogle$2.00$12.001M$5.92
GPT-5.6 TerraOpenAIOpenAI$2.00$12.001M$5.92
GPT-5.2OpenAIOpenAI$1.75$14.00400K$6.58
GPT-5.3-CodexOpenAIOpenAI$1.75$14.00400K$6.58
GPT-5.4OpenAIOpenAI$2.50$15.001M$7.40
Claude Sonnet 4.5AnthropicAnthropic$3.00$15.00200K$7.68
GPT-5.6 SolOpenAIOpenAI$4.00$20.001M$10
Claude Opus 4.5AnthropicAnthropic$5.00$25.00200K$13
Claude Opus 5AnthropicAnthropic$5.00$25.001M$13
GPT-5.5OpenAIOpenAI$5.00$30.001M$15
Claude Fable 5AnthropicAnthropic$10.00$50.001M$26
Claude Fable 5.1AnthropicAnthropic$10.00$50.001M$26
Claude Mythos 5.1AnthropicAnthropic$10.00$50.001M$26
GPT-6 AstraOpenAIOpenAI$10.00$50.001M$26
GPT-5.5 ProOpenAIOpenAI$30.00$180.001M$89

Click a column to sort. The month column assumes 700 words in, 500 words out, 20 times a day, thirty days, until you change the boxes above.

Cheapest to run

Ministral 3 14B

$0.19 a month at the default workload

Priciest to run

GPT-5.5 Pro

$89 for exactly the same work

Biggest read/write gap

Gemini 3.5 Flash-Lite

output costs 8x input, $0.30 against $2.50

One lab at a time

Same table, one lab, plus what its subscription does and doesn't cover.

AlibabaAnthropicDeepSeekGoogleMetaMistralOpenAIxAI

Why the same job costs hundreds of times more on one model than another

We wrote this up with the numbers in what 48 models cost for the same job.

Every lab sells the same two things: tokens you send and tokens you get back. Output is dearer than input everywhere, usually four or five times, because generating is the expensive half.

The rest is positioning. A small model is cheap because it's small. A flagship costs what it costs because the lab thinks you'll pay it for the hard jobs. Most people over-buy: if a cheap model already passes your own check, the expensive one is money for nothing.

What a million tokens is

Roughly 750,000 words of English, give or take. That's a long novel, twice. If you send a 700 word email and get 500 words back, you've used about 1,600 tokens, so a million tokens is around 600 of those exchanges.

Tokens aren't words, they're chunks. Common words are one token, odd ones split into several, and code and other languages split differently. Treat every number on this page as close, not exact.

What these prices leave out

Caching, batching and volume deals all cut the bill, sometimes by half or more, and none of them are in here. Subscriptions aren't either. If you use a chat app rather than the API you're paying a flat monthly fee, and none of this table applies to you.

Models with no published price sit outside the table: 4 of them, mostly open weights you run yourself and previews the lab hasn't priced.

Each model's page links the lab doc the numbers came from. Prices move, and we refetch them with the rest of the pipeline, so the stamp at the top is the honest answer to "when did you last check". Which model fits · compare two side by side · all the models · today's news