nAIvigate Barometer

The AI model market in September 2026

Name: nAIvigate Barometer September 2026
Creator: nAIvigate
License: https://creativecommons.org/licenses/by/4.0/

5 of the top 10 models were released in September 2026.

190 models tracked, 131 of them language models. Figures computed on October 1 from verified model sheets and the daily history, with no LLM involved. Free to reuse with attribution (CC BY 4.0).

Daily history available since September 7: this month's variations cover 24 days, not the full month.

Released this month

among tracked models

Median output price

7.50 $

per million tokens

Self-hostable

71 %

93 of 131 LLMs

Top of the ranking

1525Elo

Gemini 4 Argon

The month in brief

24 tracked models were released in September 2026.
An LLM via API costs a median 2.00 $ input and 7.50 $ output per million tokens. Half of the 39 models with a verified public price charge between 3.75 $ and 17.50 $ for output.
93 of 131 LLMs can run on your own infrastructure. In the cloud, 17 guarantee EU data residency and 12 offer it on request.
For the 38 proprietary LLMs, we were able to verify 88% of our criteria with the vendor on average. Enterprise requirements (Vendor private endpoint, Subprocessor list published, SCIM provisioning) are the least verified so far: collection is under way.

This month's ranking

LMArena text Elo, among tracked models.

First when computed on October 1: Gemini 4 Argon (Elo 1525). Entered the top 10: Gemini 4 Argon, Claude Opus 5.5, Claude Fable 5.1, Muse Spark 1.3. Dropped out: Gemini 3.7 Flash, Gemini 3.1 Pro, Gemini 3 Pro, GPT-5.6 Sol.

1Gemini 4 ArgonGoogle1525 released this month
2Claude Opus 4.6Anthropic1505 =
3Claude Fable 5Anthropic1505 ▼ from 1
4Claude Opus 5.5Anthropic1504 released this month
5Claude Opus 4.7Anthropic1502 ▼ from 3
6Claude Fable 5.1Anthropic1501 released this month
7Muse Spark 1.3Meta1495 released this month
8Gemini 3.8 FlashGoogle1494 released this month
9Muse Spark 1.1Meta1492 ▼ from 6
10Claude Opus 5Anthropic1491 ▼ from 5

Rated for the first time this month, after we added their Arena mapping: Kimi K3, GPT-6 Astra, Nemotron 3 Ultra 550B-A55B, GPT-5.6 Terra, GLM-5.2, Nemotron 3 Super 120B-A12B, GPT-5.6 Luna, Aya Expanse 32B, Kimi K2, GLM-5.3, Mistral Small 3.2. They do not count as top 10 entries.

What an LLM costs via API

Computed on the 39 LLMs whose public price was checked with the vendor, out of 131 tracked. Models without a public price or self-hosted only are left out.

Median input price

2.00 $per million tokens

Half of the models between 0.77 $ and 4.00 $

Median output price

7.50 $per million tokens

Half of the models between 3.75 $ and 17.50 $

Cheapest (input + output)

1GPT-6 Luna0.10 $ / 0.50 $
2Mistral Small 40.15 $ / 0.60 $
3Codestral0.30 $ / 0.90 $
4GPT-5.6 Luna0.20 $ / 1.20 $
5Mistral Large 30.50 $ / 1.50 $

This month's changes

No price change recorded on the 39 models with a verified public price (24 days of snapshots).

Corrections to our data, with no price change at the vendor: Mistral Large 3 (price aligned on 29 September with Mistral's current price list, $0.50 / $1.50; the previous figure in our database was outdated)

Can your data stay in Europe

Out of the 131 language models tracked.

On your own infrastructure

93/ 131

Published weights, self-hostable

Cloud, EU residency

17/ 131

Guarantee published by the vendor

Cloud, EU on request

12/ 131

Usually an enterprise contract

All modalities included (image, video, audio and 3D): 128 self-hostable models out of 190.

What we verified with proprietary vendors

For the 38 proprietary LLMs tracked: share of our 20 criteria (data, sovereignty, transparency, enterprise requirements, licence) for which we recorded information published by the vendor. A criterion not yet recorded is not necessarily missing at the vendor: collection of enterprise requirements is under way.

Share verified, on average

88 %

Criteria applicable to the model

More than half verified

34/ 38

The others have a provisional Trust Index

Least verified criteria so far

Vendor private endpoint22 of 38
Subprocessor list published28 of 38
SCIM provisioning28 of 38
Zero Data Retention30 of 38

Each model's Trust Index is on its sheet. We do not publish an average pitting open against proprietary models: a model you host yourself earns by construction the criteria you hold, which would make the comparison misleading. See the Trust Index

Released in September 2026 (24)

Official vendor release date, among tracked models.

Added to our catalogue this month (45)

Models released earlier, added to our tracking during the month (date added).

Method

Language models: all tracked models except image, video, music, voice and 3D generation. Prices in US dollars per million tokens, collected from vendor pricing pages; median and middle range (25th to 75th percentile) over LLMs whose input and output prices are public and verified; corrections to our own data are flagged separately and do not count as price changes. Releases: official release date within the month. Additions: older models added to the catalogue during the month. Arena ranking: LMArena text Elo as recorded when the figures are computed, compared with the ranking at the month's first daily snapshot. Models released during the month count as entries; models we rate for the first time because we just added their Arena mapping do not. A change below 10 Elo points between the first and last snapshot is not counted as a move. EU residency: guarantee published by the vendor for cloud hosting inside the European Union. Verification: applicable Trust Index criteria for which we recorded information published by the vendor; a criterion not recorded counts as missing, whether unpublished or not yet verified. Variations computed over 515 snapshots, from September 7 to September 30. No placement is sold, no figure is produced by an LLM.

Cite: "nAIvigate Barometer, September 2026", https://naivigate.com/en/barometre/2026-09

Archives: 2026-10 2026-09