Updated Estimate: AI coding agents are emitting 180,000 tonnes of carbon each year

‍

In May we published our first estimate that AI coding agents are emitting about 250,000 tonnes of CO₂e per year, using best available data and assumptions at the time. Today, we’re revising our estimate with an updated methodology and better underlying data. 

We now estimate that AI coding agents are responsible for emitting up to ~180,000 tonnes of carbon per year, with a central range of ~43,000–67,000 tonnes depending on which models are used.

The revised number was a result of two main changes: A higher count of AI commits and a smaller per-commit footprint. While we’re glad to see that the number is smaller than what we originally estimated, the forces behind these numbers are still growing quickly with increased model capability and increased developer adoption.

‍

What we found

• ~133 g of CO₂e per AI commit, down from 613 g. We updated our estimates to use the approach laid out by Oviedo et al. (Joule 2026), a recent paper by Microsoft Research that directly measures this.  We take their approach that explicitly accounts for caching but with a recent frontier model (Opus 4.8), which brings it down to ~133 g.
• ~1.35 billion AI commits a year, up from ~204 million. Only about 1.5% of public commits carry a signature, such as a commit co-authored by Claude. Most AI-written code carries no agent signature; like when a developer reviews and commits the draft, tweaks it first, pastes from a chat window, or accepts autocomplete. We account for these cases by training a classifier to recognize the patterns in the code itself, and it finds that roughly 1 in 3 public commits now involves substantial AI help. That's about 10x our initial conservative assumption. 
• ~180,000 tonnes per year today. One reasonable critique is that we are using a frontier model for impact where you could use a mix of frontier and more efficient models. We take the simpler approach, since new releases of frontier models invariably consume more tokens than their predecessors. Assuming a similar rate of model releases as prior years, what is frontier today will likely be more mid-tier in a few months. Using a recent frontier today likely overstates emissions in the short-term but may understate them in the long-term.

‍

How we estimated

We share the full detail in our updated methodology. In brief, the estimate is built from three sub-estimates and we updated two of them.

1. Carbon per commit: counting what the model writes. Tokens used for model output consume far more compute than reading them, especially with improvements to storing and accessing previous context (i.e., caching). We now use measured written and reasoning token counts from Artificial Analysis and a recent frontier model is ~133 g.
2. Commits per year: counting with a classifier instead of signatures. Our first estimate scaled up the commits that carried an AI signature and assumed we were catching about half of all agent activity. So we trained a model to detect AI-written code directly from the code and commit itself. It reaches 89% accuracy and, more importantly, is well calibrated, which is what lets us read its output as a population share. Our new analysis suggests there was 10x the agent activity we initially estimated. Applied to the public commit population and scaled to the full market, it puts the count at ~1.35 billion AI commits a year.
3. AI agent adoption rates remain steady. The share of commits involving AI rose from about 19% in January 2026 to about 36% by May; roughly four points a month with no sign of flattening. Research from Stanford’s Digital Economy Lab has found that new models tend to consume more tokens and compute for the same task. More code being produced on top of them being more energy-intensive suggests that this will continue to be a growing source of emissions. 

‍

What this means

We lowered our emissions estimate, but our original conclusions remain the same. 180,000 tonnes per year is still a lot, about the same as 39,000 gas-powered cars driven for a year. The trends are also not changing: more and more code is written with AI.

We remain committed to revising and updating our assumptions as more data is disclosed. If you have data, a critique, or a different read on the numbers, email us at feedback@cnaught.com.

Keep an eye out for a coming post where we turn to the rest of the AI ecosystem: what the published research and provider disclosures actually support for chat, image, and video generation, and what that means for anyone trying to put a number on their AI emissions.