LLM API prices, September 2026.
Prices checked , updated daily until the month ends
So far in September 2026 the cheapest model with 128k context is Llama 3.2 3B Instruct at $0.08 per million tokens from Inference, and the typical price across the market is $4, checked 29 Sept.
The month so far
Typical price
$4/M
the middle of every price sold
Models
347
from 219 providers
Makers' prices moved
18
8 cheaper · 10 dearer · from 10 Sept
New providers
16
+22 models first listed this month
Cheapest by context
The cheapest text model holding at least that much context and sold by 3 providers or more, and the typical price of those models.
32k context or more
Llama 3.2 1B Instruct
60k context
Inference
A typical one costs $1.95/M tokens, the middle of 297.
128k context or more
Llama 3.2 3B Instruct
131k context
Inference
A typical one costs $2.10/M tokens, the middle of 281.
1M context or more
Qwen3.7 Flash
1M context
Eden AI
A typical one costs $3.90/M tokens, the middle of 99.
Biggest cuts
this week's changes →The makers' own prices that ended the month somewhere other than they started.
Mistral AI · 73% cheaper · $3.20/M $12/M
OpenAI · 50% cheaper · $0.90/M $1.80/M- OpenAI · 50% cheaper · $8/M $16/M
- OpenAI · 50% cheaper · $9/M $18/M
- OpenAI · 50% cheaper · $40/M $80/M
Biggest rises
- Mistral AI · 300% dearer · $12/M $3/M
Alibaba Cloud · 170% dearer · $0.91/M $0.34/M- OpenAI · 150% dearer · $150/M $60/M
- Alibaba Cloud · 107% dearer · $2.40/M $1.16/M
Z.ai · 100% dearer · $0.95/M $0.47/M
16 new providers
First read in September, with the models each had at the last refresh.
How we count
This report is rewritten by every daily refresh until September 2026 ends, and after that it is never changed. This version was written from the refresh of 29 Sept 2026.
A move is a price that ended the month somewhere other than where it started it, in the provider's own currency, so an exchange rate moving is not a move, and a price that went down and back up again is not one either.
Our price record for this section begins on 10 Sept 2026, so this month's moves and new providers are counted from that day, and a provider already here on that day did not join.
The cheapest of a tier is a price that could lead a ranking on the day: the provider shows it as available and its prices were read in the last three days.
Every price is per million tokens at the provider's standard rate, blended 3 in : 1 out, with no cache, batch or promotional discount. The typical price is the middle of every priced row anyone sells. A free tier is a limit rather than a price, and is in none of these numbers.
A price below the maker's own with nothing published to account for it is listed on the model's page and is never a cheapest here. A tier counts only models sold by 3 providers or more, so one rarely sold variant cannot own it.
The moves are the makers' own prices, where the lab that trained the model publishes one. Resellers' prices move as well, many of them from day to day as their catalogs change; the changes page lists those as they happen.
Prices by month