2.3 mo
Traffic doubled about every 10 weeks
From 22.3T in Oct 2025 to 576T in Sep 2026 is 26x in 11 months, a doubling every 2.3 months. Any footprint that doubles that fast is out of date by the time it is reported.
Updated October 2026Case study · OpenRouterPublic traffic · Oct 2025 to Sep 2026Methodology v1.2
This update adds September 2026, with 576 trillion tokens, 52% more than in August 2026.Data fetched 8 Oct 2026, 06:47 UTC.
We metered a year of OpenRouter traffic to show what an unmeasured AI footprint actually looks like.
A single gateway now routes 576 trillion tokens a month, 127.1 kt of CO₂e, up 26x in a year, and none of it appears in a Scope 3 disclosure. Your own usage curve has the same shape.
1,852T
tokens routed, Oct 2025 to Sep 2026
408.9 kt
CO₂e over the twelve months, methodology v1.2
26x
growth in monthly tokens over the year
70%
of Sep 2026 traffic went to six providers
What this study shows
Between Oct 2025 and Sep 2026, OpenRouter’s public rankings recorded 1.9 quadrillion tokens routed through one gateway, with monthly volume growing 26x. Run through CrbonFree methodology v1.2, that traffic amounts to about 408.9 kt of CO₂e for the year, concentrated in six providers, and reported by no one at this granularity.
Fetched 8 Oct 2026, 06:47 UTC from the product's public endpoint, which mirrors openrouter.ai/rankings as of 2026-10-07. Page last checked 5 October 2026. Every figure on this page is computed from the dataset below.
Use the numbers
The monthly series, the computed CO₂e with bounds and the provider share are published as JSON so a spreadsheet, a script or an agent can check them.
Download the datasetCite as: Crbon Labs, “OpenRouter public traffic, metered with CrbonFree methodology v1.2”, data from OpenRouter as of 2026-10-07 (fetched 8 oct 2026, 06:47 utc).
Five things the year shows
34%
Compound growth, month on month
The average month added 34% on the one before. Five months, Feb, Mar, Jun, Aug and Sep 2026, each grew by more than half.
70%
Six providers carry 70% of the traffic
DeepSeek, Z-AI, OpenAI, Tencent, Xiaomi, Google together took 70% of Sep 2026. Concentration like that means a handful of provider choices decide most of the footprint.
88,882
A year of this traffic is a mid-sized city’s cars
408.9 kt of CO₂e is what about 88,882 passenger cars emit in a year, or roughly 408,855 economy round trips between New York and London.
Sources: US EPA, 4.6 t per typical passenger vehicle per year; ICAO carbon calculator, about 1 t per round trip
±28%
Every figure comes with its band
The Sep 2026 figure of 127.1 kt sits in a band from 91.2 kt to 163.1 kt. The band is stated on purpose: a number without one cannot be audited.
Tokens routed per month
22.3T in Oct 2025 to 576T in Sep 2026. The curve is public, but nobody publishes the footprint underneath it. Hover a month for its figures.
1.9 quadrillion tokens over twelve months
OctNovDecJanFebMarAprMayJunJulAugSep
Month by month
Tokens are as published by OpenRouter, and the CO₂e comes from methodology v1.2 at the medium tier, assuming a 75/25 input to output split. Every row can be reproduced from the factors below.
| Month | Tokens | Share of peak | vs prior | CO₂e est. |
|---|---|---|---|---|
| Oct 2025 | 22.3T | – | 4.9 kt | |
| Nov 2025 | 26.6T | +19% | 5.9 kt | |
| Dec 2025 | 26T | -2% | 5.7 kt | |
| Jan 2026 | 31.7T | +22% | 7.0 kt | |
| Feb 2026 | 49.8T | +57% | 11.0 kt | |
| Mar 2026 | 83.8T | +68% | 18.5 kt | |
| Apr 2026 | 97.1T | +16% | 21.4 kt | |
| May 2026 | 123T | +27% | 27.1 kt | |
| Jun 2026 | 188T | +53% | 41.6 kt | |
| Jul 2026 | 248T | +32% | 54.8 kt | |
| Aug 2026 | 379T | +53% | 83.7 kt | |
| Sep 2026 | 576T | +52% | 127.1 kt | |
| Twelve months | 1,852T | 26x | 408.9 kt |
Who the tokens went to
Share of Sep 2026 traffic by model provider, from OpenRouter’s public rankings, with the CO₂e each share implies at the month’s total of 127.1 kt.
Share of Sep 2026’s tokens, %
ProviderShare of tokensTokensCO₂e
DeepSeek21.2%122T tokens26.9 kt CO₂e
Z-AI13.3%76.8T tokens16.9 kt CO₂e
OpenAI13.0%74.8T tokens16.5 kt CO₂e
Tencent12.5%71.7T tokens15.9 kt CO₂e
Xiaomi5.4%31.4T tokens6.9 kt CO₂e
Google4.6%26.6T tokens5.8 kt CO₂e
Other providers30.0%173T tokens38.1 kt CO₂e
Crbon Labs Inc. (2026)
What each update added
This page is rewritten from OpenRouter’s data each time a month completes. Every line here is recomputed from today’s dataset, so it shows what that update reads now.
Updated October 2026
Added September 2026, with 576 trillion tokens, 52% more than in August 2026. The study then covered twelve months, 1.9 quadrillion tokens and 408.9 kt of CO₂e.
Published September 2026
The first edition covered eleven months, October 2025 to August 2026, with 1.3 quadrillion tokens and 281.7 kt of CO₂e.
How 576 trillion tokens in September 2026 became 127.1 kt of CO₂e.
This is methodology v1.2, the same three layers the product applies to a customer’s own usage, and it adopts Watershed’s open framework for measuring emissions from AI usage. It is written out here so an auditor can redo the arithmetic.
Per trillion tokens, medium tier
221 t CO₂e
About 459 MWh at the facility meter after host power, utilisation and PUE.
- 1
Active energy
Each token is charged joules for the work it causes: 0.3 J per input token read, 0.5 J per output token generated, 0.03 J when served from cache. Aggregate counts carry no cache data, so nothing is credited to caching here. The 75/25 split is the methodology's worked-example assumption for gateway traffic.
- 2
Facility operation
Accelerators do not run alone. Host power adds 18%, clusters average 30% utilisation, and the data centre's overhead is a PUE of 1.2. Grid intensity is 0.42 kg CO₂e per kWh, a global average for the medium tier.
- 3
Full lifecycle
Hardware manufacturing adds 0.02 kg and model training 0.008 kg per million tokens, amortised over the fleet's serving life. This is the figure reported everywhere on this page.
- ±
Uncertainty
IPCC Tier 1 root sum of squares of 20% activity and 20% factor uncertainty: ±28.3%. For Sep 2026 that is a band of 91.2 kt to 163.1 kt around 127.1 kt.
- ×
What is excluded
The public dataset is aggregate, so there is no per-model tiering. Network transfer, end-user devices and the gateway’s own routing infrastructure are left out, and the product reports water separately. Every one of these exclusions pushes the number lower, so the figure on this page is a floor.
What a Scope 3 disclosure would need that this page cannot give
A public dataset gets you a shape and a scale. An auditor asks for more. Each row pairs the question with what this page has and what CrbonFree adds when the usage is yours.
The auditor asksThis page hasWith your own usage
Which models and providers produced it?
Provider share for one month, from the public rankings.
Per model and per provider, per call, in the dashboardWhich factors, and which version?
Methodology v1.2, medium tier, stated below.
The full methodology with every factorHow sure are you?
A band of plus or minus 28.3% on every figure.
The same band on every figure your own usage producesWho on your team used it?
Not in a public dataset.
The CLI meters coding agents per person and per modelCan an agent or a script check the number?
The dataset behind this page, as JSON.
Seven read-only MCP tools and a typed SDK
Four terms this page leans on
- Token
- The unit a language model reads and writes, roughly three quarters of an English word. Providers bill by the token, and so does carbon accounting for AI, because the energy a request uses scales with the tokens it processes.
- Scope 3
- The GHG Protocol category for emissions a company causes but does not own, such as purchased services. AI inference bought from a provider sits here for the buyer, which is why an unmeasured AI footprint is a Scope 3 gap.
- PUE
- Power usage effectiveness: total data-centre energy divided by the energy the computing equipment itself uses. A PUE of 1.2 means cooling and power distribution add 20% on top of the servers.
- Uncertainty band
- The range a figure is expected to fall in, given how well its inputs are known. This study combines a 20% activity uncertainty and a 20% factor uncertainty by root sum of squares, giving plus or minus 28.3%.
Questions
Questions about this study.
Short answers. The methodology page has the factors and the arithmetic.
Read the methodologyNo. OpenRouter publishes aggregate token counts in its public rankings, and this page applies the CrbonFree methodology to those counts. The carbon figures are ours, not OpenRouter’s, and the assumptions behind them are listed on this page.
They belong to the organisations sending the requests, as purchased services under Scope 3, and to the providers running the models, as the energy their data centres draw. The point of the study is that neither side reports them at this granularity today.
Each figure carries a band of plus or minus 28.3%, the root sum of squares of a 20% activity uncertainty and a 20% factor uncertainty. The public dataset is aggregate, so no per-model factors are applied, which pushes the estimate down rather than up.
The public counts do not separate input from output, and output tokens cost more energy to generate than input tokens cost to read. The methodology’s worked example for gateway traffic assumes three quarters input and one quarter output, and this page uses the same assumption.
Yes. Connect a provider with a read-only key, install the CLI for coding agents or switch on the browser extension, and the dashboard reports your usage per model and per provider with the same methodology and the same uncertainty band.
Sources and data
- Dataset
- OpenRouter public rankings, openrouter.ai/rankings, as of 2026-10-07. Fetched 8 Oct 2026, 06:47 UTC.
- Methodology
- CrbonFree methodology v1.2, medium tier, 75/25 input to output split, adopting Watershed’s open framework.
- Equivalents
- US EPA, 4.6 t CO₂ per typical passenger vehicle per year. ICAO carbon calculator, about 1 t per economy round trip New York to London.
- Download
- The dataset as JSON, monthly tokens with CO₂e and bounds, provider share, methodology and source. Published 2026-09-03, last checked 5 October 2026.
See your own number instead of ours.
However your team already runs AI, that is how you connect. Paste a read-only provider key, drop in an SDK, add the MCP server, install the CLI or switch on the browser extension, and the first measured tokens reach your dashboard within seconds.
Create your free account
Free plan · first tokens in seconds