Coroid reports estimated carbon emissions for the model inference your work consumes, and offers a Low carbon AI profile that biases routing toward lower impact.
What the estimate covers
Operational inference only — the energy used running models to do your work, combined with the carbon intensity of the grid serving the region the model ran in.
What it does not cover
Being precise about the boundary matters more than the number:
- Not model training. The very large one-off cost of creating a model is excluded.
- Not embodied carbon. Manufacturing the hardware is excluded.
- Not your own infrastructure. Your CI, your test runs, your deployments.
- Not a carbon-neutrality claim. Nothing here asserts offsetting.
It is an estimate of a bounded, real quantity. Treat it as a comparative signal — useful for telling whether one configuration is heavier than another — rather than as an audited absolute.
The Low carbon profile
Prefers smaller models running on lower-intensity grids. Both levers matter: a smaller model does less work per request, and the same work emits differently depending on where it runs.
The trade-off is real. Smaller models are less capable across the whole pipeline, so Low carbon suits routine work better than architecturally significant work. Choose it per piece of work rather than as a blanket default — see AI profiles.
Reducing impact without changing profiles
The largest reduction available to most teams is not model selection — it is avoiding wasted work:
- Reduce rework. A task that runs twice emits twice. Tight specifications are the lever. See Specifications.
- Scope tasks tightly. Broad tasks explore more to establish their own boundaries.
- Keep context lean. Context is most of the tokens.
These are the same actions that reduce cost, because cost and emissions are both driven by tokens.
Where to see it
Model profiles show carbon characteristics alongside speed, quality and cost, so you can weigh all four when assigning a profile.