What is the carbon footprint of a startup's own AI use? Our October numbers
By Guillermo Llopis, CTO at SOMA AI
Page updated
What is the carbon footprint of a startup's own AI use? Our October numbers
Short answer. We estimate that the AI use we recorded at SOMA AI in October 2026 caused about 316 kg CO2e, with a plausible range of 113 to 481 kg. The same run gives about 382 kWh of electricity and 775 litres of water. 98% of the emissions came from coding agents, not from chat or from the AI inside our product. These are estimates, and they cover only part of the company.
What did we count?
We counted the AI use we could read from logs and consoles for October 2026.
- Claude Code on one laptop. Token counts come from the local session logs, split into fresh input, cache reads and output, per model (Fable 5.1, Opus 5, Opus 5.5, Sonnet 5.5 and Haiku 4.5).
- Claude Code on our background server. The queue that runs automated engineering work, including subagents. It runs on a subscription plan, not on our API keys, so it does not overlap with the API line below.
- Anthropic API, SOMA keys only. Read from the Anthropic console. We included only keys used for SOMA work and left out other projects that share the same console organisation.
In October the laptop logs add up to about 7.6 billion tokens and the server to about 4.8 billion. The SOMA API keys used about 119 million.
What is our estimate in kg CO2e?
We uploaded the October usage as a CSV to our AI footprint calculator, which applies SOMA's published method. It returns 316 kg CO2e for the month, with a range of 113 to 481 kg.
As a cross-check we ran the same October data through the factor library inside the SOMA product. It gives about 248 kg.
| Source | October (kg CO2e) |
|---|---|
| Claude Code, laptop | 166 |
| Claude Code, server | 76 |
| Anthropic API, SOMA keys | 5.2 |
| Product inference, GreenPT | 0.8 |
| Total | about 248 |
The two figures differ because of the grid assumption. The library places Anthropic inference in the US East region, while the calculator uses the global average grid of 483 g CO2e per kWh. Both land inside the same range. Anthropic does not offer EU inference, so we could not assume a cleaner region. The region default should be the same in every SOMA tool, and we will align it.
Why do coding agents dominate?
Coding agents re-read their whole context on every turn. These cache reads dominate by volume. In October the laptop alone logged about 7.3 billion cache-read tokens against roughly 25 million output tokens.
Our published method weights a cache-read token at 0.05 of an output token, with a range of 0.02 to 0.10. Moving the weight across that range shifts the coding-agent lines between roughly 0.5 and 1.8 times the central estimate.
The lesson for anyone estimating coding-agent use is to count cache reads. Leaving them out can understate the footprint of an agent-heavy team by around an order of magnitude.
What is not in these numbers?
This is not the footprint of the whole company. The following are still missing.
- Claude Code on two other machines we use.
- Our SEO citation harness on OpenAI, Perplexity and Gemini.
- Chat, phone and desktop app use by the rest of the team.
- AI features inside other software, such as Gemini in Workspace, Canva or Notion.
- A marketing agency's Midjourney campaign in October, estimated separately at about 3 kg CO2e.
Product inference through GreenPT is also missing from the calculator figure, because the calculator cannot take it yet. In the library run it adds 0.8 kg.
How can your company do the same?
Start with the tools that already log usage. Coding agents and API consoles give token counts per model, which beat guessing from seat numbers. Then separate cache reads from fresh input, because that choice can move your result by a large factor. Apply an emissions factor per model and region, and keep the assumptions written down.
Our guide explains the full process in how to measure the environmental impact of the AI you use. If you want to try it with an export of your own, the AI footprint calculator accepts a CSV of tools, models and token counts.
What did we learn?
The biggest lesson is that where the footprint sits depends on how you work. For us it is not the chatbot or the product, it is the coding agents that run all day. We also publish the assumptions next to the number, because the same data can give different results under different grid defaults.
FAQ
- What is SOMA AI's own AI carbon footprint?
- We estimate about 316 kg CO2e for October 2026, with a range of 113 to 481 kg. Almost all of it, 98%, comes from coding agents.
- Does this cover the whole company?
- No. It covers Claude Code on one laptop and on our server, plus SOMA's Anthropic API keys. Other machines, the rest of the team's chat use and AI inside other software are not included yet.
- Why do you say estimate and not measurement?
- Token counts are recorded from logs, but the energy and emissions per token are modelled from public factors. We do not measure the electricity of the data centres we use.
- Why do cache reads matter so much?
- Coding agents re-read their context on every turn, so cache-read tokens are the bulk of the volume. Even at a low weight of 0.05 of an output token they dominate the total.