How I calculated 2.2 million AI tokens

The headline estimate is 2.2 million AI tokens, rounded from 2,232,000. It describes an assumed monthly volume of text processing associated with selected work. With the task-by-task model choices below, its estimated API cost is US$18.27 at prices checked on 1 November 2026. This is the unadjusted reference cost; ability and equivalence remain unassessed for this profile. It is a tool budget, not the value of a person.

This companion gives the workings behind Am I only worth 2.2 million AI tokens?. Use the calculator to replace the work volumes, choose models for individual tasks and compare subscription fees with API spending.

Start with outputs you can count

My roles include Chief AI Officer, board member for ten companies, physicist, business mentor, writer, graduate lecturer, regional-association leader, and AI engineer and scientist. I have more than 15 years of experience across these fields. The quantities, text lengths and model assignments below are planning assumptions, not a record of a particular month or a claim about which products I used.

Count each deliverable once. A board paper that also supports an executive briefing is not automatically two separate outputs. Meetings, experiments and live teaching cannot be reproduced merely by generating text about them.

Planning assumptions. Scroll the table horizontally on narrow screens. Row costs are rounded after calculation.

Separate the archive, training and future work

The 2,232,000-token reference is a monthly processing allowance for selected future outputs. It is not the token equivalent of a career. To estimate accumulated output, count the actual permitted documents or code from defined periods, remove duplicates and record gaps. If a measured average is extrapolated over working years, state the comparable periods and changes in role or output volume. A tokenised archive describes recorded text, not the full expertise developed while producing it.

Supervised fine-tuning adapts a model using labelled examples. Its dataset is a selected subset, not automatically every past output. Count the prepared training data and repeated training passes separately from the tokens processed when you later use the model; apply the supported model’s training tariff rather than the serving rates in this worksheet. Preparation, evaluation and deployment also have costs. This calculator does not quote a fine-tuning job or assume that every listed frontier model supports it.

Assess a tuned system on representative unseen cases kept separate from training. Google’s fine-tuning guidance warns that overlapping training and test examples can inflate performance. Repeating familiar answers is insufficient evidence of future replacement. A prospective comparison must cover the required service, new cases, changing conditions, ongoing review, retained knowledge and responsibility. The calculator’s acceptance record is a limited task-cost projection; it cannot establish that an entire role can be removed.

Experience defines the comparison standard

The calculator accepts relevant years of experience for each role. My aggregate experience is more than 15 years across the fields listed above; I have not assigned that tenure to every role. Years give context for the human work standard and the difficult cases to include. The calculator does not infer competence or multiply tokens by tenure.

Test ability, output and reasoning on the same cases

Choose a task scope the tool can actually attempt with the available information, permissions and tools. Record cases tested and cases completed. For completed cases, count outputs meeting the experienced-human comparison standard and approaches meeting predefined reasoning criteria. Record how many passed both checks together. Keep the reviewer, criteria and workflow/version reference with the record.

Output equivalence concerns correctness, coverage and usefulness under the same acceptance standard. Reasoning equivalence here concerns observable decision quality: relevant evidence, valid assumptions, constraints, exceptions, alternatives and appropriate deferral. It does not establish identical internal thinking. Anthropic’s reasoning-faithfulness research cautions against treating a model’s explanation as a reliable account of its internal process. Billed reasoning tokens are a cost input, not an equivalence score.

Use representative cases, including difficult cases, and assess the person’s work under the same conditions. NIST’s AI Risk Management Framework calls for documented test sets representative of expected use. Set criteria before inspecting answers; where feasible, have reviewers assess outputs without knowing which system produced them. This worksheet records your assessment; it does not conduct a formal statistical equivalence test or independently verify the submitted evidence.

How accepted-output costs are estimated

Joint acceptance rate equals cases passing both output and reasoning checks divided by all cases tested. The denominator includes cases the tool failed to complete. The calculator divides each task’s reference token allowance and API cost by that observed rate to project the resources for the requested number of accepted outputs. It does not multiply separate completion, output and reasoning percentages, which may describe overlapping failures.

As a worked arithmetic example, suppose every task had five jointly accepted cases out of ten tested cases. Holding average workflow cost and future case mix fixed would double the reference allowance to 4,464,000 tokens and the task-mix API estimate to about US$36.54. These counts are an example, not measured performance of the listed models. The headline’s 2.2 million remains the unadjusted reference workload.

Zero jointly accepted cases provide no finite accepted-output estimate. Missing records, a partially feasible scope, a blocked task or a tested-model mismatch withhold the whole-workload comparison. Changes to the model, work standard or workflow assumptions require confirmation again. Repeating a workflow may reproduce the same failure, so the projection is not a promise that retries will succeed. Small or unrepresentative samples cannot establish dependable performance on future cases.

API and subscription comparisons can use either the reference workload or the sample-conditioned accepted-output projection. Confirmed unused API grants apply only to the matching provider. Enter human hours at the same accepted-output standard, including supervision and rework. Test setup and continuing review costs are additional unless included in the entered hours or other costs; the time-value result remains withheld until the comparison record is complete for every requested task.

See the token arithmetic

The selected outputs total 93,000 final word-equivalents. Using a rough English-text conversion of four tokens per three words and two full drafting passes gives 248,000 generated output tokens. An input allowance of eight tokens for every generated token adds 1,984,000 input tokens. Together they make 2,232,000, rounded to 2.2 million in the headline.

Across the 46 selected outputs, that is an average processing allowance of about 48,522 reference tokens per output, including input and drafting. Individual tasks have different lengths. The input allowance represents instructions, source material and conversational context across the drafting process. The word-equivalent for code is an uncalibrated placeholder. Actual tokenizers differ, so these are reference counts rather than measured counts for each provider. Extra internal reasoning, retries beyond the two passes, searches, tools, media and integration can increase the bill. The baseline does not add extra hidden reasoning; change that assumption when the chosen model bills it.

Use demanding models for the tasks that need them

The model mix reserves Claude Fable 5.1 for board-paper challenge and difficult technical work, GPT-6 Astra for strategic work, and Gemini 4 Argon for scientific analysis. Sonnet, Sol, Flash and Opus cover the other selected outputs. A model’s price or position in a product range does not establish that it is the best choice for a task; test the result and change the assignment.

Each row is priced as input tokens divided by one million times the input rate, plus output tokens divided by one million times the output rate. Adding the rows gives US$18.27. Sending this entire reference workload to Fable would instead cost US$32.24; sending it all to GPT-6.1 Sol would cost US$6.45. These are cost comparisons using the same assumed counts, not equivalent-quality results.

API rates behind the estimate

USD per million tokens, standard uncached text API rates. No batch, caching or speed premiums. Sources: OpenAI, Claude, Gemini.

Model	Input USD / million	Billed output USD / million	Rate valid through
OpenAI GPT-6.1 Sol	2.00	10.00	No stated promotion end
OpenAI GPT-6 Astra	10.00	50.00	No stated promotion end
Claude Sonnet 5.5	2.00	10.00	No stated promotion end
Claude Opus 5.5	4.00	20.00	No stated promotion end
Claude Fable 5.1	10.00	50.00	No stated promotion end
Gemini 3.8 Flash	0.75	3.75	2026-12-31
Gemini 4 Argon	2.00	10.00	2026-11-30

Gemini 4 Argon changes from USD2/10 to USD4/20 per million input/output tokens on 1 December 2026. Gemini 3.8 Flash’s listed promotional rate lasts through 31 December. Hidden reasoning can increase billed output. See how token counts work and use model-specific measured counts before relying on a budget.

What a subscription changes

An app subscription pays for access, features and usage subject to the plan’s limits. It does not establish that this workload fits within a fixed monthly token allowance. The following are US monthly web prices before tax, without introductory trials or annual discounts, checked on 1 November 2026.

For the OpenAI and Claude API benefits, eligibility includes a seven-day wait for new qualifying subscriptions and linking or claiming the benefit. Grants expire at the end of the applicable billing cycle and do not roll over. Use only the unused amount confirmed in your account. Claude’s app usage credits and its Platform API grants are separate balances; a grant is not transferable to another provider.

Compare the dollar totals without double-counting

API only: the task mix costs about US$18.27 before the excluded costs. Keeping ChatGPT Plus, Claude Pro and Google AI Pro costs US$59.99 in plan fees; that does not prove the entire task mix fits their app allowances. If you keep those three subscriptions and also process the complete workload through separately billed APIs, the combined estimate is US$78.26.

A higher-plan example is ChatGPT Pro 100, Claude Max and Google AI Ultra at US$99.99: US$299.99 in fees. If the qualifying OpenAI and Claude API grants have been claimed and enough remains unused, they cover their own US$6.10 and US$10.75 portions of this task mix. The Google API portion remains about US$1.42, giving US$301.41 in fees plus uncovered API use. This example does not recommend buying all three plans or establish app capacity.

Make an estimate for your own work

Choose the outputs you need. Count tokens for a permitted sample of typical outputs in each task and use their average as the starting allowance; the table above uses assumed average text lengths. Enter the quantities and lengths, then assign a model to each task. Check a sample with the selected model’s tokenizer and include repeated context, reasoning and correction. Compare the time and cost of reaching the same usable standard. The calculator keeps those assumptions editable, shows both reference costs and evidence-conditioned accepted-output projections, and lets you enter confirmed unused subscription API benefits.

Before using the estimate in a staffing decision, test whether the service remains dependable and how essential knowledge will be retained.

Prepared with AI assistance and editorial review.