Short answer
Four levers move the bill and the carbon estimate in the same direction: a smaller model class, reasoning only where it is needed, cached input, and shorter prompts and outputs. Two levers do not. Batch processing halves the price but SAEF does not credit it with lower carbon, and a region with a data-residency premium can cost more while sitting on a cleaner or a dirtier grid. Compare cost and carbon per task and not per token, because tokenizers differ between models.
Every carbon number here is a SAEF estimate, not a measurement. For the organisation-wide view that includes seat tools, see the company-use page. This page covers the API side.
Which levers move cost and carbon, and in which direction?
| Lever | Cost effect (provider pricing, checked 21 November 2026) | Carbon effect (SAEF estimate) |
|---|---|---|
| Smaller model / routing | Lower per-token price. | Lower, up to 13 times from class C to class A. |
| Reasoning effort | Lower, because reasoning tokens are billed as output tokens. | Lower, fewer output tokens. At k = 25 the estimate is 12 times the plain exchange. |
| Cached input | Lower, cached input costs roughly a tenth of fresh input. | Lower, a cached token weighs 0.05 against 0.3. |
| Shorter outputs and prompts | Lower, output costs several times input. | Lower. Halving the output gives 1.3 times lower. |
| Batch | Lower, 50% off. | No change. SAEF gives batch no credit. |
| Region | Same, or about 10% higher with residency endpoints. | Depends on the grid, up to 12 times apart. |
Does a smaller model or routing cut both?
Yes. Smaller models carry lower per-token prices at the major providers, and SAEF groups models into three size classes with the estimate for one exchange at 0.01 g for class A, 0.036 g for class B and 0.13 g for class C in US East - N. Virginia. Routing classification, extraction and short drafting to a smaller class and keeping the largest class for hard tasks is the biggest single lever on both sides. Test quality on your own tasks first. The named-model comparison shows how SAEF maps models to classes.
Does limiting reasoning effort cut both?
Yes. Reasoning tokens are billed as output tokens, which OpenAI states in its reasoning guide, and Google lists output prices including thinking tokens on its pricing page (both checked 21 November 2026). Anthropic reports the thinking share of the billed output tokens in the usage object of its extended thinking API (checked 21 November 2026). Output costs several times input, 5x on every Claude model, 4x to 8x across OpenAI's listed models and 5x on Gemini 4, so reasoning tokens are expensive tokens.
On the carbon side, more output tokens mean a higher SAEF estimate. Multiplying the output of one class B exchange by k, an author assumption of 25 with a range of 10 to 50, gives 0.44 g, against 0.036 g without reasoning, with 0.19 g and 0.87 g at the ends of the range. In API use you do not need the assumption, because the billed output tokens already include the reasoning. See the reasoning models page.
Does cached input cut both?
Yes. Anthropic, OpenAI, Google and Mistral all bill cached input tokens at a discount to fresh input. A tenth of the input price is typical, and newer models go lower (OpenAI, Google, Mistral, checked 21 November 2026). On Claude, cache writes cost extra, 1.25 times the input price for a five-minute cache and 2 times for a one-hour cache, and pay back after one or two reads (Anthropic pricing, checked 21 November 2026). Google adds an hourly storage charge on explicit caches.
SAEF weights an uncached input token at 0.3 and a cached input token at 0.05, against 1 for an output token. As a scenario we chose to show direction, not a typical rate, reading half of the 928 input tokens from cache gives 0.029 g for a class B exchange, 1.2 times lower than the plain 0.036 g. Training emissions are counted per billed token, cached or not, so the carbon falls by less than the price does.
Do shorter prompts and outputs cut both?
Yes. Every token is billed and every token is in the SAEF estimate, and output tokens cost several times more than input tokens. Cutting the output of a class B exchange from 282 to 141 tokens gives 0.027 g, 1.3 times lower, again as a labelled scenario. Ask for concise answers, set a maximum output length and trim pasted context.
Does batch processing cut emissions?
It cuts the price, not the SAEF estimate. All four providers sell an asynchronous batch mode at half the standard price. Anthropic's pricing page gives a 50 percent discount on input and output tokens, every batch price on OpenAI's page is half of standard, Google lists a 50% cost reduction for its Batch API on its pricing page, and Mistral says batch processing reduces the price by 50% on its pricing page (all checked 21 November 2026). A discount for accepting a delay is a price, not evidence that the provider uses less energy, so SAEF does not credit batch with lower carbon. Use batch to cut cost on work that can wait, and report its carbon at the standard estimate.
Where can cost and carbon pull in opposite directions?
In the region. Keeping inference in one region costs more. Anthropic's inference_geo: "us" adds 1.1 times to every token category on Claude 4.6 and later, and Bedrock and Google Cloud regional endpoints carry a 10 percent premium over global ones for Claude 4.5 and later (Anthropic pricing, checked 21 November 2026). OpenAI charges a 10% uplift on regional processing endpoints for models released on or after 5 March 2026 (OpenAI pricing, checked 21 November 2026).
Carbon depends on the grid of the region, not on the premium. Across the 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. A residency endpoint in a clean-grid region can therefore cost 10% more and cut the estimate, and one in a carbon-intensive region can do the opposite. Where the provider does not say where a model runs, ask and record the answer, because the region is an input to the estimate. The grid factors are location-based annual averages, not statements about any provider's electricity contracts.
Why compare cost and carbon per task and not per token?
Tokenizers differ. Claude 4.7 and later use a tokenizer that produces about 30% more tokens for the same text, so per-token prices are not comparable across tokenizer generations (Anthropic pricing, checked 21 November 2026). Both the bill and the SAEF estimate follow billed tokens, so a model with a lower price per token can still cost more and estimate higher on the same document. Run the same task set through each candidate, then compare cost per task and estimated kg CO2e per task.
How do I track cost and carbon side by side?
Log input, cached input and output tokens per call with the model and region. Cost is list price times tokens. Carbon is the same tokens times a SAEF factor. The saef package does the second step. Install it with pip install saef.
import saef
usage = {"input_tokens": 1_500_000, "output_tokens": 400_000}
r = saef.estimate(**usage, model="gpt-4o-mini", region="eu-west-1")
print(f"central {r.kg_co2e:.4f} kg CO2e, high case {r.kg_co2e_high_case:.4f} kg CO2e")
print(f"water {r.water_ml:.1f} mL")It has integrations for LiteLLM, LangChain, LlamaIndex, Haystack and the OpenAI and Anthropic SDKs, listed on the developer tools page. LiteLLM, Langfuse and Helicone already record cost per call, so the carbon estimate can sit in the same record. The step-by-step logging method is in how to track emissions from AI API calls.
For a first estimate without code, use the free AI footprint calculator. The full SOMA AI calculator imports OpenAI, Anthropic, LiteLLM, Helicone and OpenRouter usage exports and shows list-price cost next to the carbon estimate, with saving actions in euros and kg CO2e. The factors are in the factor table.
What does this page not claim?
- That any carbon figure is a measurement. They are SAEF estimates from tokens, model class and grid.
- That batch processing lowers carbon. It lowers price, and SAEF gives it no credit.
- That a smaller model or a lower reasoning effort keeps your quality. Test it on your tasks.
- Current prices. Provider prices change, so this page gives effects and cites each provider page as checked on 21 November 2026, with no per-model price table.
- That the reasoning multiplier is observed. It is an author assumption, shown with its range.
- That a provider runs a model in a given region. Regions are inputs to the dataset, and the grid factors are location-based averages.
Short answers
- What is the fastest way to cut both the cost and the emissions of LLM API usage?
- Route routine work to a smaller model, keep reasoning modes for tasks that need them, and cache the fixed part of long prompts. Each of these removes tokens or moves them to a cheaper model, so it lowers the bill and the SAEF estimate together. SOMA AI estimates, from its open SAEF emission factors, that the model class changes the estimate for one exchange by about 13 times from class A to class C.
- Does the Batch API lower the carbon footprint?
- Not in the SAEF estimate. Anthropic, OpenAI, Google and Mistral all price batch processing at half the standard rate, but that is a price for accepting a delay and not evidence that the work uses less energy. SAEF does not credit batch with lower carbon, so batch is a cost lever only.
- Can the cheapest region be the highest-carbon region?
- Yes. Keeping inference in a named region usually costs about 10% more, and the carbon of that region depends on its grid. SOMA AI estimates, from its open SAEF emission factors, that a class B exchange comes to 0.008 g on the Quebec (Canada) grid and 0.095 g on the Bahrain grid, so the choice can raise cost and lower carbon, or the reverse.
- How do I put cost and carbon in the same report?
- Log input tokens, cached tokens and output tokens per call with the model and region, price them with the provider's list prices, and estimate carbon from the same counts with the saef package or the SAEF CSV. Gateways such as LiteLLM, Langfuse and Helicone already record cost per call, so the carbon estimate can sit in the same record. Label the carbon figure an estimate.
Keep reading
- How can a company reduce the cost and carbon footprint of its AI use without losing quality? The organisation-wide version, including seat tools
- How can I track emissions from AI API calls for carbon accounting? Logging tokens and totalling kg CO2e by month
- How much more energy do reasoning models use than standard chat? Benchmarks and the multiplier behind the reasoning lever
- Developer tools. The saef package, integrations and open dataset
