Market share for AI models is usually reported from surveys of companies, which show who bought what and little about how much work the models do. By token count, OpenRouter routed 575.9 trillion tokens in September 2026, 26 times its October 2025 volume, and four Chinese labs carried 52% of those tokens, led by DeepSeek at 21.2%. This post charts twelve months of OpenRouter’s public token counts, shows which providers and models gained share, and prices the traffic through Crbon Labs methodology v1.2 at about 127,000 tonnes of CO2e for September 2026.
A survey asks companies which vendor they use, so a company running a billion tokens a day and one running a thousand count the same. Tokens are the unit the models are paid in, and OpenRouter publishes them, per model, every day.
OpenRouter publishes tokens per model, for one part of the market
OpenRouter is a gateway that takes a developer’s request at one endpoint and sends it to whichever provider the developer chose, so OpenRouter can see which models get used, and it publishes the counts on its rankings page and through a dataset behind it. We sum that dataset by month and by provider, and we run every month through Crbon Labs methodology v1.2, the method the product uses to turn tokens into CO2e. The sums in this post are as of 5 October 2026 and cover the twelve complete months from October 2025 to September 2026.
OpenRouter is a sample rather than the whole market. Traffic that goes straight to OpenAI’s or Anthropic’s own API doesn’t pass through OpenRouter, and neither does anything inside ChatGPT, Gemini or Copilot. The developers who route through a gateway are more price-sensitive than average, and some of the largest applications on it are chat and role-play products. That skews the shares towards cheap and free models. OpenRouter’s rankings are still the largest public dataset that counts tokens per model, and a skew towards cheap models does not change which way the volume and the shares are moving.
OpenRouter’s monthly tokens grew 26x in twelve months
- Tokens OpenRouter routed in the month
Crbon Labs Inc. (2026)
Monthly tokens went from 22.3 trillion in October 2025 to 575.9 trillion in September 2026. That is 26x in eleven steps, a doubling about every 10 weeks, or 34% growth a month. September 2026 added 196.5 trillion tokens to August 2026’s 379.4 trillion tokens, a 52% increase in one month and more than the gateway’s whole volume of 188.4 trillion tokens in June 2026. The 12 complete months add up to 1.85 quadrillion tokens. On our factors that is about 409,000 tonnes of CO2e, which is what about 89,000 cars emit in a year, from one gateway.
OpenRouter’s tokens in September 2026 alone came to about 127,000 tonnes of CO2e, with a band of 91,000 to 163,000 tonnes of CO2e, and about 264 gigawatt-hours of electricity at the data centre meter once host power, cluster utilisation and cooling are counted.
Four Chinese labs carry 52% of OpenRouter’s tokens, and Google’s share fell to 4.6%
- DeepSeek21.2%
- Z-AI13.3%
- OpenAI13%
- Tencent12.5%
- Xiaomi5.4%
- Google4.6%
- Other30%
Crbon Labs Inc. (2026)
DeepSeek is the largest provider on the gateway, with 21.2% of September 2026’s tokens. Its growth came in one step. DeepSeek’s share sat between 5% and 8% from October 2025 to April 2026, then V4 Flash arrived in April 2026 and the share went to 15.1% in May 2026 and 22.4% in August 2026. Three builds of its Flash model, V4 Flash from April and July 2026 and V4.1 Flash from September 2026, carried 19.3% of September 2026’s tokens.
Z-AI doubled its share in a month, from 6.7% in August 2026 to 13.3% in September 2026. The increase came from GLM 5.3 Flash, a model dated 26 August 2026 that became the most used model on the gateway within five weeks, at 9.9% of tokens. Tencent moved the same way. HY4 Preview, dated 27 August 2026, took 9.7% of September 2026’s tokens. With HY3 at 2.7% of tokens, Tencent holds 12.5%. Xiaomi went the other way, from 14.8% in July 2026 to 5.4% in September 2026. Together DeepSeek, Z-AI, Tencent and Xiaomi held 52.4% of September 2026’s tokens, up from 47.9% in August 2026 and 9.1% in October 2025.
OpenAI’s share is the highest it has been in the twelve months. It was 10.1% in October 2025, fell to 5.3% in July 2026 and reached 13.0% in September 2026, with GPT 5.6 Luna at 8.9% of tokens. The gateway’s total grew 26x over the year, so OpenAI’s token volume grew about 33x.
Google was the largest of the six providers until April 2026, peaking at 23.7% of tokens in January 2026, and it ended September 2026 at 4.6%, a fifth of that peak and the largest fall of the six. Google’s token volume still grew 3.5x over those eight months, because the gateway grew 18x. Anthropic’s Claude 4.5 Sonnet was the second most used model on OpenRouter in October 2025, at 10.1% of tokens, and no Claude model is among September 2026’s fifteen most used. That is a statement about who routes Claude through OpenRouter, not about Claude’s use in general.
The six largest providers carried 37.8% of OpenRouter’s tokens in October 2025 and 70.0% in September 2026, and the rest of the gateway went from 62.2% of tokens in October 2025 to 30.0% in September 2026.
Counted in tokens on OpenRouter, more than half of the AI work in September 2026 went to DeepSeek, Z-AI, Tencent and Xiaomi, and the most used model was GLM 5.3 Flash, in a version dated 26 August 2026. A market share measured once a year, or by asking companies which vendor they pay, would have missed the four Chinese labs passing half of OpenRouter’s tokens and a most used model that was five weeks old.
Six of September 2026’s fifteen most used models were dated 26 August 2026 or later
| # | Model | Tokens | Share, % | CO2e |
|---|---|---|---|---|
| 1 | GLM 5.3 FlashZ-AI | 56.9T | 9.9% | 12.6 kt |
| 2 | HY4 PreviewTencent | 56.1T | 9.7% | 12.3 kt |
| 3 | DeepSeek V4.1 FlashDeepSeek | 51.4T | 8.9% | 11.3 kt |
| 4 | GPT 5.6 LunaOpenAI | 51.3T | 8.9% | 11.3 kt |
| 5 | DeepSeek V4 FlashDeepSeek | 42.5T | 7.4% | 9.4 kt |
| 6 | MiMo V2.5Xiaomi | 19.9T | 3.5% | 4.4 kt |
| 7 | Nemotron 3 Ultra 550B A55BNVIDIAfree | 19.2T | 3.3% | 4.2 kt |
| 8 | DeepSeek V4 FlashDeepSeek | 17.3T | 3% | 3.8 kt |
| 9 | HY3Tencent | 15.6T | 2.7% | 3.4 kt |
| 10 | GLM 5.3Z-AI | 12.3T | 2.1% | 2.7 kt |
| 11 | MiMo V2.6 FlashXiaomi | 9.8T | 1.7% | 2.2 kt |
| 12 | Gemini 3.8 FlashGoogle | 9T | 1.6% | 2.0 kt |
| 13 | GPT 5.6 SolOpenAI | 7.8T | 1.4% | 1.8 kt |
| 14 | GLM 5.2Z-AI | 7.5T | 1.3% | 1.7 kt |
| 15 | Muse Spark 1.3 ContributorMeta | 7.3T | 1.3% | 1.7 kt |
Crbon Labs Inc. (2026)
Six of the fifteen most used models in September 2026 carry a version date of 26 August 2026 or later in OpenRouter’s listing. They are GLM 5.3 Flash, HY4 Preview, DeepSeek V4.1 Flash, MiMo V2.6 Flash, Gemini 3.8 Flash and Meta’s Muse Spark 1.3 Contributor, and together they carried 33.1% of September 2026’s tokens. First place changed hands just as often. Eight different models were the most used on OpenRouter in the twelve months, and September 2026’s leader, GLM 5.3 Flash, held 2.1% of tokens in August 2026.
OpenRouter’s dataset treats each build of a model separately, so the April and July 2026 builds of DeepSeek V4 Flash and the September 2026 build of V4.1 Flash are three rows, and each newer build took over from the last within weeks of its version date. NVIDIA’s Nemotron 3 Ultra, which costs nothing to call, ranks seventh on OpenRouter with 3.3% of September 2026’s tokens.
OpenRouter’s tokens in September 2026 come to about 127,000 tonnes of CO2e
Crbon Labs methodology v1.2, which adopts Watershed’s open framework for measuring emissions from AI usage, turns tokens into CO2e in three layers: the energy the chips use for the tokens themselves, the energy the facility uses to keep those chips running and cool, and a small allowance for building the hardware and training the model. On medium-tier factors, with the aggregate split we use for any token count we have not metered (three quarters input, one quarter output, nothing cached), a trillion tokens comes to about 221 tonnes of CO2e. Multiply by September 2026’s 575.9 trillion tokens and you get about 127,000 tonnes of CO2e.
The uncertainty band deserves as much attention as the total. Two 20% uncertainties, one on activity and one on the factors, combine to plus or minus 28%, so September 2026 sits somewhere between 91,000 and 163,000 tonnes of CO2e. The table’s per-model column shares the month’s total out in proportion to tokens, because the public dataset says nothing about which of those tokens were input, output or cached, and nothing about the hardware behind each model. A metered account replaces the assumed split of three quarters input and one quarter output, and the assumed hardware, with a count per call, which is the point of measuring rather than estimating.
Four levers cut a month’s CO2e roughly in half
None of the four levers changes what you build; each one changes what a token costs in energy and in money, and those two costs move together.
Cache what repeats. System prompts, tool definitions and the documents you paste into every call are read again on every call unless the provider serves them from cache. On our factors a cached input token costs a tenth of a fresh one. Serving half of your input from cache cuts a month's CO2e by about 25%; serving four fifths cuts it by about 40%. OpenAI caches repeated prefixes of 1,024 tokens or more on its own; Anthropic needs a cache marker in the request. Both charge a tenth of the input price for a cache read, which is why this is the lever that gets pulled.
Send routine work to a smaller model. Classification, extraction, routing and short replies rarely need a flagship. A small-tier token costs about a third of a medium-tier token on our factors, so moving half of the traffic to a small model cuts the total by about 28%, and moving all of it cuts it by more than half. On OpenRouter the market has already moved: three of September 2026’s five most used models are Flash variants.
Ask for shorter answers. Output tokens cost more energy than input tokens, about 0.5 joules against 0.3 joules on the medium tier, because each one is generated in turn. Bringing output from a quarter of tokens down to 15%, with tighter instructions or a lower maximum length, saves about 5%, which is a small saving and a free one.
Mind the grid. The same tokens in a data centre with a PUE of 1.09, the ratio of the facility’s power to its computers’ power, on a grid of 0.345 kilograms of CO2e per kWh emit 22% less than in one with a PUE of 1.2 on a grid of 0.42 kilograms of CO2e per kWh. Region choice is a lever for anyone on a cloud provider that offers it, and it works on every token regardless of the other three.
Pull the first three together, half the input cached, half the traffic on a small model, output at 15%, and a month's CO2e roughly halves.
1B
A billion tokens a month is a mid-sized team on the API. OpenRouter routes a few hundred trillion.
0.3 J per input token, 0.5 J per output token, 0.03 J when cached.
0%
System prompts, documents and tool definitions that repeat between calls.
25%
Output tokens cost more energy than input tokens. Shorter answers move this down.
0%
Classification, extraction and short replies rarely need the flagship.
CO2e a month, after the levers
221 kg
from 221 kg before, plus or minus 28.3 percent on both
The calculator’s starting total is an estimate on the aggregate split, and a metered account replaces it with a count, which changes the total and, more often than people expect, changes which lever matters most. The calculator also stops at the emissions that remain. Retiring verified credits against the measured total funds real removals, and on our platform that is a separate step which doesn’t touch the measured total, so the report stays honest and the climate action sits on its own line.
What 26x growth on OpenRouter means for a team that buys AI through an API
OpenRouter’s tokens grew 26x in twelve months, to 575.9 trillion in September 2026 and about 127,000 tonnes of CO2e for that month, and the models carrying them changed faster than the volume did. A team that buys AI through an API is on the same curve at a smaller scale. The share it sends to each model moves month by month, the builds change under it, and its emissions follow its tokens, so an estimate made last quarter describes a smaller and different workload. An emissions total you can defend needs the count per call, the factor per model tier and the method written down, and the count is the one part that cannot be reconstructed afterwards. CrbonFree meters the tokens per call, per model and per provider as they are used and converts them with Crbon Labs methodology v1.2, so a team cuts its emissions against a measured total and retires credits against what is left.
Sources
- 01OpenRouter, LLM Rankings
The public rankings page behind the dataset: tokens per model, updated continuously.
- 02OpenRouter API, rankings-daily dataset
Tokens per model per day for the trailing 360 days; the monthly token counts in this post are sums of it, last refreshed on 5 October 2026. October 2025 is the first complete month in the dataset.
- 03CrbonFree methodology v1.2
The tier factors, the three layers and the uncertainty band used for every figure on this page.
- 04Bistline et al., “Estimating GHG Emissions from AI Use: Framework for Corporate-Level Measurement”, Watershed, August 2026
The open framework methodology v1.2 adopts: token volumes, energy per token, PUE and grid.
- 05CrbonFree, OpenRouter case study
The same twelve months with the method worked through in full, month by month.
- 06Anthropic, Prompt caching
How cached prefixes are reused across calls; cache reads are billed at a tenth of the input price.
- 07OpenAI, Prompt caching
Automatic caching of repeated prompt prefixes of 1,024 tokens or more.
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.