Building with AI? Count Thinking Tokens Before You Scale

Got an AI idea? Give it one useful job first.

A small, representative task gives you something concrete to judge: does the answer help, is it fast enough, and what does the complete run cost?

Watch the 39-second video

Start with a result you can recognise

Pick one example from the workflow you want to build. Write down what a good answer must contain before you run it. Compare quality, speed and total cost together. A low-priced answer that needs repeated retries may not be the best fit for your task.

Count the tokens you cannot see

The visible answer is only part of the budgeting picture. Include input tokens and billable output, including reasoning or thinking tokens where the model charges for them. Check how your selected model reports usage: if reasoning tokens are already included in the output total, count them once.

Then include applicable platform fees and every additional call or retry in the workflow. Use the rates and usage records for the model you actually test, rather than estimating cost from the length of the final answer.

Save this first-task checklist

• Choose one representative input.
• Define the result you need.
• Try a model and check the answer against that definition.
• Record response time and the complete billed usage.
• Include thinking tokens where applicable, without double-counting them.
• Include applicable fees, extra calls and retries.
• Use what you learn to decide the next test.

One example is a starting point, not proof that a model will work for every case. Test a broader set before relying on it in production.

Turn the idea into a first test

You do not need the whole application finished to learn which model fits your task. Start with a useful example, measure the result, and build from there.

Create your unoblox account and try your first task.

The accompanying video uses AI-generated narration.