Can token counts measure AI carbon emissions?

No. A token count records model activity, not electricity use or carbon emissions. Converting tokens into CO₂e requires assumptions about the model, hardware, runtime, data centre, electricity mix, and system boundaries. Without that evidence, the result is an estimate whose uncertainty should be stated.

Tokenforest uses token counts for a separate purpose: a clear rule that funds tree planting as AI usage accumulates. It does not present weighted tokens as measured emissions.

What a token count actually tells you

A token is a unit of text processed or produced by a model. Provider usage records can split tokens into input, output, cached input, cache writes, and reasoning.

Those counts are useful for comparing activity within a consistent setup. They can also support cost checks and usage trends. They do not record power draw, elapsed accelerator time, cooling overhead, or the carbon intensity of the electricity used.

Why the same token count can use different amounts of energy

At least six variables sit between tokens and carbon:

1. Model architecture and size. Different models perform different amounts of computation per token.
2. Hardware. Accelerator type, memory, and server generation affect power and throughput.
3. Utilisation and batching. A request may share hardware efficiently or run with spare capacity.
4. Context and caching. Reading cached context is not the same operation as computing new model states.
5. Data-centre overhead. Cooling and power delivery add energy beyond the accelerator itself.
6. Electricity source. The carbon intensity of the grid varies by place and time.

A credible estimate needs data or documented assumptions for these variables. A credible measurement needs metered energy and a defensible method for assigning it to the workload.

How Tokenforest weights usage

Tokenforest first separates overlapping provider totals, so cached input and reasoning are not counted twice. It then applies a stable product weighting:

Token class	Weight
Input	1×
Output	3×
Cache write	1.25×
Cache read	0.02×
Thinking	3×

The formula is:

weighted tokens =
  input
  + 3 × output
  + 1.25 × cache write
  + 0.02 × cache read
  + 3 × thinking

One tree of progress equals 50,000,000 weighted tokens. This threshold decides when tree funding is due. It is not an emissions factor.

What the 84 kg figure means

Tokenforest may show 84 kg CO₂e as lifetime sequestration context for a planted tree. It describes a long-term tree estimate, not the emissions of the AI sessions that funded it.

Planting a tree also does not erase a specific model request. Survival, growth, time, species, location, and land management all affect real sequestration. That is why Tokenforest treats planting as a contribution tied to usage, not a one-for-one carbon offset.

What responsible AI carbon reporting should include

If you publish an AI emissions estimate, state:

• whether energy was measured or modelled;
• the hardware and model covered;
• the start and end of the system boundary;
• the location and time basis for electricity carbon intensity;
• how shared infrastructure was allocated;
• the uncertainty range and known omissions;
• whether any tree or credit figure is separate from the emissions estimate.

Tokenforest's methodology page publishes the current product rules and their limits. The CLI's local report supplies an auditable activity count, while the environmental claim remains bounded to what the available data supports.

A useful role for tokens

Tokens are still a practical trigger for action. They are available in local coding-agent records, can be separated by class, and can be counted without collecting source code or prompts.

That makes them suitable for a predictable funding rule. It does not turn them into an electricity meter. Keeping those two ideas separate makes the result easier to inspect and harder to overstate.

To inspect your own records, use the local guides for Claude Code or Codex.