AI units

Every call is counted in AI units, whatever the model. Models that cost more to run use more units for the same tokens.

How a call’s units are worked out

Each model has a number of units for every million tokens it reads, and for every million it writes. See the Input units and Output units in Models.

A call’s units are:

(input tokens × input units + output tokens × output units) ÷ 1,000,000

rounded up to a whole unit.

An example

A call to openai/gpt-oss-120b that reads 1,000 tokens and writes 500:

(1,000 × 20,000 + 500 × 79,000) ÷ 1,000,000 = 59.5, so 60 units

Every answer says the units it used, in usage.units.

When the numbers are estimated

If a stream stops before the model reports its tokens, we count the tokens from what was sent and written, and the units from those. This is rare.

What a unit costs

What units cost depends on your plan. A project’s AI page in the console shows the units it has used, and your organisation’s estimate shows what they cost.