Work, tokens and judgement
What could tokens help you produce?
Describe your roles and monthly outputs. Estimate text-processing costs, then compare the human time needed to reach a result you can use.
Your expertise is not converted into tokens. Years of experience define the comparison standard alongside tested ability, output and reasoning criteria; they do not apply an invented token multiplier. A lower API bill cannot establish that a role can be replaced.
Past work, fine-tuning and future work
This estimates a future month of selected work. Relevant years describe the human comparison standard; the 2.2 million reference is not a career-output total or a measure of accumulated expertise.
Career output needs a separate count of actual documents or code across defined periods. Fine-tuning adapts a model using selected examples; training passes and training tariffs are separate from future API usage. This tool does not quote training costs or assume every listed model supports fine-tuning.
Future replacement requires evidence on unseen work, changing conditions, ongoing review and responsibilities. Reproducing past documents cannot settle that question. Keep evaluation cases separate from training examples. See the method and source guidance.
Your planning estimate
Estimated monthly token use
Reference text API cost
Human time difference
Time-value comparison after entered costs
This subtracts API and entered other costs from the value you assign to the time difference. It uses the sample-conditioned accepted-output API cost once every requested task has a complete comparison record. Enter human hours at that same accepted-output standard, including rework. This is a capacity scenario, not cash savings, salary value or demonstrated return.
What a monetary budget buys
Ability, output and reasoning comparison
Scroll the table horizontally on narrow screens. A record must match the costed model and the current workflow. Reasoning equivalence means meeting observable decision criteria, not proving the model thinks like a person. A longer explanation or a higher reasoning-token bill is insufficient.
| Role | Relevant experience | Evidence status | Joint passes / cases tested | Monthly projected reference tokens | Monthly projected API USD |
|---|
The projection divides the reference workload and API cost by the observed joint acceptance rate. It assumes comparable future cases and workflow costs; it is not a formal statistical equivalence test or a guarantee that retries solve persistent failures. No measured ability or equivalence data is prefilled. Subscription totals below use the selected cost basis: reference workload or sample-conditioned accepted-output projection. These remain API spending scenarios, not app-capacity guarantees. Include test setup and continuing review in human hours or other costs.
Costs by task
Model assignments are choices to test, not evidence of capability. Use demanding models for work that needs them and change the choices after checking results.
| Role | Model | Estimated tokens | USD |
|---|
Compare dated API prices
| Model | USD / million input tokens | USD / million billed output tokens | Monthly scenario | Source |
|---|
These prices do not establish equivalent capability. The estimate uses standard first-party, uncached, non-batch text API prices. Gemini’s listed rate is promotional. Price and access changes require a fresh provider check.
Compare subscriptions and API spending
Optional US monthly plans, before tax. These fees do not buy a fixed 2.2 million-token allowance. Select only plans you intend to keep. Enter unused API benefits only after claiming them; benefits stay with their provider and expire.
ChatGPT plan source · API benefit terms
Claude plan source · API benefit terms
Plan fees plus uncovered API usage
App limits and model access still apply. Claude Pro uses extra usage credits for Fable; Max and Ultra include Fable up to half the weekly allowance. Google Flow credits are for creative features, not the text API estimate here. See the method and current plan sources.
Inputs stay in this page and are not stored between visits. Download the scenario if you want to keep it. This tool makes no AI API calls.
Test the output before claiming a benefit
Choose a real task and decide what a usable result must contain. Compare work with and without AI at that standard, including checking and correction time. Keep the person approving the result able to question it, and record what went wrong.
Read Adrian’s essay and the method → · Measure broader AI value →