How can a company cut the carbon footprint of its AI use without losing quality?

Use a smaller model class for routine tasks, switch reasoning modes on only where the task needs them, and prefer lower-carbon serving regions where you can choose. In the SAEF dataset, the model class changes the estimate for one exchange by a factor of about 13 from class A to class C, and the region by a factor of about 12 from the lowest to the highest cloud-region grid. Start by estimating your current use, because the levers only pay off where the volume is.

These are estimates, not measurements. Every number below is one exchange of 928 input and 282 output tokens, the token assumption of the SOMA AI Emission Factors (SAEF), and none of it says your quality will hold, because that depends on the task. The SAEF method is accepted for publication in Sustainable Production and Consumption.

There are six levers, listed here in the order the dataset can size them.

• Model size class. Sized below from the class factors.
• Serving region. Sized below from the regional grid factors.
• Reasoning modes. Sized below with SOMA's reasoning multiplier, an author assumption.
• Cached input and shorter prompts and outputs. Sized below as two labelled scenarios.
• Seat right-sizing. Qualitative, because inference follows usage and not licences.
• Estimating first. Free tools for the baseline.

Yes, in the SAEF estimates. SAEF groups models into three size classes, from small (A) to frontier (C), each the mean of open-weight models measured on the ML.ENERGY Leaderboard. The same exchange in US East - N. Virginia estimates as follows.

So routing routine work such as classification, short summaries or drafting from a template to a smaller class moves the estimate by up to 13 times per exchange. Whether the smaller model does the task as well is something you test on your own tasks. See the named-model comparison for how SAEF maps each model to a class, including its limits.

Yes, because electricity is most of the estimate on a carbon-intensive grid and the grid differs by region. Across the AWS, Azure and Google Cloud regions in the dataset, a class B exchange estimates 0.008 g on the Quebec (Canada) grid and 0.095 g on the Bahrain grid, about 12 times apart for identical work. Inside the EU, Sweden estimates 0.009 g and Poland 0.070 g, about 7.4 times. Hardware, construction and amortised training do not depend on the grid, so they set a floor under the low end. The grid factors are location-based annual averages, so a lower figure reflects the grid a region sits on and not any renewable electricity a provider buys.

You can act on this only where your provider lets you choose a region, for example through a regional endpoint or a data residency setting, and only if the region meets your legal and latency needs. Where you cannot choose, ask the provider where the model runs and record the answer, because the region is an input to your estimate. The country comparison ranks the grids behind this.

Yes. A reasoning mode generates more output tokens for the same question. Billed tokens already include those reasoning tokens, so for API use the lever shows up directly in the usage export. For seat tools that export no token counts, SOMA assumes a reasoning message writes k times the output tokens, an author assumption of 25 with a range of 10 to 50, explained on the reasoning models page. The table applies k to one exchange to illustrate the size of the effect.

For a whole workforce the effect is smaller, because most messages do not use a reasoning mode. SOMA's blended uplift on output tokens is 2.20, from a reasoning share of 5% of messages (see the Scope 3 guide). The lever is to keep that share low by using reasoning for multi-step analysis and not for lookups, reformatting or short answers.

Yes, and the dataset has a factor for the first. SAEF weights an uncached input token at 0.3 and a cached input token at 0.05, against 1 for an output token, so reusing a long fixed prompt costs little. The amortised training term is counted per billed token, cached or not, so the cache row falls by less than the weights alone suggest. The two rows below are scenarios we chose to show the direction, not typical rates. They use a class B exchange in US East - N. Virginia, against the plain 0.036 g.

Beyond these two scenarios we give no figure. How much a company can shorten prompts or reuse a cache depends on the workflow. Asking for concise answers, trimming pasted context and keeping a stable prompt prefix are the practical steps. If your usage export reports cached tokens separately, the estimate takes the credit, and if it does not, the estimate runs high.

Only through usage. An AI licence that nobody opens sends no requests, so it has little inference footprint of its own, and cancelling it mainly saves cost. The footprint follows messages and tokens, which is why SOMA estimates seat tools from active users or message counts and not from licence counts. Right-sizing still helps you estimate better, because it shows who actually uses the tool and how heavily. The method is on the seat-licensed tools page.

Estimate first, so you know which lever matters for your own mix of tools. The free AI footprint calculator gives a first estimate from your tools and volumes without registration. Once you have a baseline, the general Reduction Planner turns a Scope 1, 2 and 3 footprint into reduction actions ranked by impact and effort. The class mapping is on the methodology page and the full table is in the factor table.

• That a smaller class or a non-reasoning mode gives the same quality. That depends on the task, so test it.
• That any provider runs a model in a given region or offers a lower-carbon option. Regions here are inputs to the dataset and not statements about providers.
• That the figures are measurements. They are estimates per exchange with the dataset's token assumptions, so real exchanges will differ.
• That the grid factors reflect renewable contracts. They are location-based annual averages.
• That the reasoning multiplier is observed. It is an author assumption, shown with its range.
• That the cache and output scenarios are typical. They only show direction.

• How much more energy do reasoning models use than standard chat? Benchmarks and the multiplier behind the reasoning lever
• Where does an AI query emit the least? The carbon footprint of AI by country. The grids behind the region lever
• Which AI model has the lowest carbon footprint? GPT, Claude, Gemini, Llama, DeepSeek and Mistral compared. How named models map to size classes