15 Sep 2026
Right-sized model = biggest lever on tokens and emissions
Since earning the Green Software Practitioner certification, I've been reading model launch announcements differently.
Every model launch this year came with two numbers: price per token, and benchmark scores. Almost none came with a third — how much energy it uses.
In the last two weeks, five frontier models shipped: Claude Fable 5.1, Claude Mythos 5.1, GPT-6 Astra, Gemini 3.8 Flash, Meta Muse Spark 1.3. None of them published a per-token energy or carbon figure.
That gap matters, because the carbon cost of using an LLM runs through a chain:
- The model you pick sets how many tokens a task takes.
- Tokens set energy use, in watt-hours.
- Energy, multiplied by the carbon intensity of the grid powering the datacentre, sets emissions in CO₂e.
The vendor only tells you about link one. And the spread is wide: Microsoft's own 2026 study (published in Joule) puts an efficient, at-scale query at about 0.3 Wh. The University of Rhode Island's estimate for a standard GPT-5 response averages ~18 Wh. Switch on extended "reasoning" for that same query and it can run 5–10× higher — into the hundreds of watt-hours. Same task category, a three-order-of-magnitude spread, and no vendor tells you which end of it you're on.
So what actually pulls both the bill and the emissions down at the same time? Using the right-sized model for the job. Route the ~80% of easy requests to a small model and keep the frontier model for the hard 20%. Same output, far fewer tokens — the same target Anthropic's 75%-cheaper cache reads and Google's 88%-fewer-token video processing are chasing from the other side.
One caveat, which a follow-up post takes on properly: cheaper, more efficient tokens tend to get used more (the rebound effect), so efficiency per task is necessary but not sufficient.
On a model-selection matrix I'm building, I added a column the vendors don't give you: estimated kgCO₂e per million tokens, with the assumptions written down. Having to estimate it is the point.
Sources: Microsoft, "Energy Use of AI Inference, Efficiency Pathways, and Test-Time Scaling" (2026, Joule); University of Rhode Island GPT-5 energy estimate, as reported by Tom's Hardware and Data Center Dynamics.
Originally posted on LinkedIn ↗