OpenRouter publishes the largest public record of tokens per model in the AI industry, and the volume of tokens and the models used are changing so rapidly that we cannot rely on last quarter’s data to tell the story for this quarter. OpenRouter’s model rankings for September 2026 add up to 575.9 trillion tokens, 52% more than in August 2026, with a most used model that was five weeks old and a third of the month’s tokens on model versions dated 26 August 2026 or later. This post summarises the September 2026 rankings in comparison to August 2026, prices the month through Crbon Labs methodology v1.2 at about 127,000 tonnes of CO2e, and explains what that pace means for a company that reports its AI emissions.
OpenRouter is a gateway that routes developers’ requests to AI models from many labs, and its rankings are a public leaderboard of those models by the tokens they process, a token being the unit of text a model reads and writes. We sum those counts by month, and as of our refresh on 5 October 2026, September 2026 is the largest of the twelve complete months in the dataset. Our post on LLM market share in tokens covers the full year, and this post covers what changed in one month.
OpenRouter routed 575.9 trillion tokens in September 2026, 52% more than in August 2026
OpenRouter routed 575.9 trillion tokens in September 2026, against 379.4 trillion tokens in August 2026. The 196.5 trillion tokens added in one month are more than the 188.4 trillion tokens of the whole month of June 2026. Between August and September 2026 the daily average went from 12.2 trillion tokens a day to 19.2 trillion tokens a day.
- Average tokens a day over the complete month
- Average tokens a day, 1 to 4 October 202623.2
Crbon Labs Inc. (2026)
The first 4 days of October 2026 averaged 23.2 trillion tokens a day, 21% above September 2026’s daily average. At this pace we forecast that OpenRouter will route approximately 718 trillion tokens in October 2026.
Over the twelve complete months from October 2025 to September 2026, OpenRouter’s monthly tokens grew 26x, from 22.3 trillion tokens to 575.9 trillion tokens, a doubling about every 10 weeks. Five of the eleven month-to-month changes in OpenRouter’s tokens were increases of more than 50%, and August 2026 and September 2026 were two of them in a row.
The most used model on OpenRouter in September 2026 was five weeks old
OpenRouter lists each build of a model with a version date, and a model’s share is its percentage of all the tokens OpenRouter routed in a month. GLM 5.3 Flash, from the Chinese lab Z-AI, was the most used model on OpenRouter in September 2026, with 9.9% of the month’s tokens. Its version date is 26 August 2026, and it held 2.1% of tokens in August 2026. Tencent’s HY4 Preview, dated 27 August 2026, was second with 9.7% of tokens. DeepSeek’s V4.1 Flash, dated 10 September 2026, was third with 8.9% of tokens in the three weeks after its version date. The other seven of the ten most used models were OpenAI’s GPT 5.6 Luna, two builds of DeepSeek’s V4 Flash, Xiaomi’s MiMo V2.5, NVIDIA’s Nemotron 3 Ultra, Tencent’s HY3 and Z-AI’s GLM 5.3.
| Model | Lab | Version date | Sep 2026 share | Aug 2026 share |
|---|---|---|---|---|
| GLM 5.3 Flash | Z-AI | 26 Aug 2026 | 9.9% | 2.1% |
| HY4 Preview | Tencent | 27 Aug 2026 | 9.7% | under 1.7% |
| DeepSeek V4.1 Flash | DeepSeek | 10 Sep 2026 | 8.9% | under 1.7% |
| GPT 5.6 Luna | OpenAI | 9 Jul 2026 | 8.9% | 6.6% |
| DeepSeek V4 Flash | DeepSeek | 31 Jul 2026 | 7.4% | 12.3% |
| MiMo V2.5 | Xiaomi | 22 Apr 2026 | 3.5% | 7.9% |
| Nemotron 3 Ultra | NVIDIA | 4 Jun 2026 | 3.3% | 4.2% |
| DeepSeek V4 Flash | DeepSeek | 23 Apr 2026 | 3.0% | 6.3% |
| HY3 | Tencent | 6 Jul 2026 | 2.7% | 9.2% |
| GLM 5.3 | Z-AI | 16 Aug 2026 | 2.1% | under 1.7% |
Six of the fifteen most used models in September 2026 carry a version date of 26 August 2026 or later, and together those six carried 33.1% of the month’s tokens. Four of August 2026’s five most used models lost share in September 2026. The July 2026 build of DeepSeek V4 Flash went from 12.3% of tokens to 7.4%, Tencent’s HY3 from 9.2% to 2.7%, Xiaomi’s MiMo V2.5 from 7.9% to 3.5%, and the April 2026 build of DeepSeek V4 Flash from 6.3% to 3.0%. OpenAI’s GPT 5.6 Luna gained, from 6.6% of tokens to 8.9%.
Eight different models were the most used on OpenRouter in a month between October 2025 and September 2026, and no model was first for more than three of those twelve months.
Chinese AI models from four labs carried 52.4% of OpenRouter’s tokens in September 2026
DeepSeek, Z-AI, Tencent and Xiaomi are four Chinese companies that build AI models, and OpenRouter routes requests to models from all four.
- DeepSeek
- Z-AI
- Tencent
- Xiaomi
Crbon Labs Inc. (2026)
DeepSeek, Z-AI, Tencent and Xiaomi together carried 52.4% of OpenRouter’s tokens in September 2026, against 47.9% in August 2026 and 17.4% in April 2026. Z-AI’s share doubled in one month, from 6.7% of tokens to 13.3%, and Tencent’s went from 10.3% to 12.5%. DeepSeek stayed the largest lab on the gateway at 21.2% of tokens, and Xiaomi fell from 8.5% to 5.4%.
OpenAI’s share increased as well, from 10.5% of tokens in August 2026 to 13.0% in September 2026, its highest share in the twelve months. Google’s share fell from 7.4% of tokens to 4.6%, and no Anthropic model was among the fifteen most used. OpenRouter is a gateway that price-sensitive developers route through, so each lab’s share of OpenRouter’s tokens describes OpenRouter’s traffic and says nothing about ChatGPT, Gemini or Claude used directly.
OpenRouter’s September 2026 tokens come to about 127,000 tonnes of CO2e, against about 84,000 tonnes of CO2e for August 2026
| Month | Tokens routed | CO2e | Uncertainty band | Electricity at the data centre meter |
|---|---|---|---|---|
| June 2026 | 188.4 trillion | 41,600 t | 29,800 to 53,300 t | 86 GWh |
| July 2026 | 248.4 trillion | 54,800 t | 39,300 to 70,300 t | 114 GWh |
| August 2026 | 379.4 trillion | 83,700 t | 60,100 to 107,400 t | 174 GWh |
| September 2026 | 575.9 trillion | 127,100 t | 91,200 to 163,100 t | 264 GWh |
Crbon Labs methodology v1.2, which adopts Watershed’s open framework for measuring emissions from AI usage, prices a token count in three layers: the electricity the chips use, the electricity the data centre uses around them, and an allowance for building the hardware and training the model. OpenRouter’s counts do not separate input tokens from output tokens, so we apply the methodology’s split for aggregate counts, three quarters input and one quarter output with nothing cached, on the factors of its medium tier, the tier of models such as GPT-4o and Claude Sonnet. On that basis a trillion tokens come to about 221 tonnes of CO2e, and September 2026’s 575.9 trillion tokens come to about 127,000 tonnes of CO2e, with a band of 91,000 to 163,000 tonnes of CO2e.
September 2026’s total is about 4,200 tonnes of CO2e a day, and about 264 gigawatt-hours of electricity for the month at the data centre meter. August 2026 came to about 84,000 tonnes of CO2e, so one month added about 43,000 tonnes of CO2e to the gateway’s monthly total.
The estimate applies one medium-tier factor to every token, because OpenRouter’s rankings name the model and do not state which chips served it or which grid powered them. Our post on tokens per watt prices the same billion tokens on three chips and three grids. A metered account replaces the assumed split of three quarters input and one quarter output with counts of input, output and cached tokens for every call.
An emissions estimate built on OpenRouter’s June 2026 token count covers a third of its September 2026 emissions
On OpenRouter, the tokens routed in September 2026 were 3.1 times the tokens routed in June 2026. An emissions estimate built on June 2026’s token count comes to about 41,600 tonnes of CO2e, a third of the 127,100 tonnes of CO2e the same method gives for September 2026. A company whose AI usage grows as OpenRouter’s traffic did, and which counts its tokens once a quarter, reports a third of its current monthly emissions by the time the next count is due.
A third of September 2026’s tokens ran on model versions that were five weeks old or less at the end of the month, so a list of emission factors keyed to model names is out of date within weeks of being written. Crbon Labs methodology v1.2 assigns each model to a tier, such as small, medium, large or reasoning, and applies that tier’s published factor, so a model released last week has a factor on the day it is first metered.
OpenRouter’s rankings also leave out where the tokens were served. Watershed’s framework asks providers to disclose energy per token by model and by region, and until providers publish that, a token count per model is the firmest input a buyer of AI has.
What OpenRouter’s September 2026 rankings mean for your own AI emissions
OpenRouter’s rankings for September 2026 add up to 575.9 trillion tokens and about 127,000 tonnes of CO2e, 52% more tokens than in August 2026, with a most used model that was five weeks old. For a team that buys AI through an API, and whose usage grows as OpenRouter’s traffic did between June 2026 and September 2026, a token count from the prior quarter understates the current month’s workload by two thirds, and a count from the prior month understates it by one third. A list of emission factors written for last quarter’s models does not cover the models doing the work now. CrbonFree meters tokens per call, per model and per provider as they are used, and prices each month with Crbon Labs methodology v1.2, which integrates Watershed’s open framework for measuring emissions from AI usage. A team that tracks its emissions with CrbonFree can manage and report accurate token usage by model, daily and month over month.
Sources
- 01OpenRouter, LLM Rankings
Tokens per model, updated continuously. The version date quoted for a model in this post is the date in its permanent identifier on OpenRouter.
- 02OpenRouter API, rankings-daily dataset
Tokens per model per day; the monthly and daily token counts in this post are sums of it, last refreshed on 5 October 2026 with coverage to 4 October 2026.
- 03CrbonFree methodology v1.2
The tier factors, the three layers, the split for aggregate counts and the uncertainty band behind every CO2e total in this post.
- 04Bistline et al., “Estimating GHG Emissions from AI Use: Framework for Corporate-Level Measurement”, Watershed, August 2026
The open framework methodology v1.2 adopts; it asks providers to disclose energy per token by model and region.
- 05CrbonFree, OpenRouter case study
The twelve months from October 2025 to September 2026, month by month, with the dataset as JSON.
- 06CrbonFree, “LLM market share in tokens: a year of OpenRouter”
Provider shares for every month of the twelve and the fifteen most used models.
Figures attributed to CrbonFree come from methodology v1.2, published in full with every factor and formula. Read the methodology.
About the author
Cory Bergh
Cory leads Crbon Labs, which originates its own climate projects and builds CrbonFree, the platform that measures the footprint of AI usage per token. He was previously VP of Technology and Innovation in the energy industry and holds a BComm, an MBA and the CFA.
Share this post
Send it to whoever asked the question.