Start with an outcome you can count. An AI project budget becomes easier to check when every cost is attached to a defined task, an expected number of executions or a monthly expense. Keep assumptions alongside the estimate so you can replace guesses with observations.
Reading-time estimate: 200 words per minute for the article and source names, rounded up. Navigation and related cards are excluded.
Define a task before counting requests
A task is the outcome you initiate: review one document, draft one response or classify one enquiry. A request is one interaction with a model. A document review might extract facts, draft findings and check the draft. Counting only the final request misses the preceding work.
Write down what counts as success. A draft needing substantial manual repair may count as a failed automated task even when the model returns a response. Use the same definition when measuring success and comparing scenarios.
Estimate volume from a concrete workload
Use expected work arrivals, not a model's technical request limit. Count representative tasks over a period that includes busy and quiet days, and record how you scaled that count to a month. Keep seasonal peaks separate.
Track executions for each step. If 100 initiated tasks produce 120 executions, the average is 1.2. Include retries and repeated work. Do not apply a second retry multiplier to an average that already includes them.
Find representative token quantities
Start with provider-reported input and output usage from representative past requests. Average comparable workloads, including busy and quiet periods, rather than relying on one unusually short or long example. Historical usage measures those requests; it cannot guarantee the length or billing of future ones.
When usage records are unavailable, the optional sample-text counters can help estimate a prompt or an existing response. They count supplied text with cl100k_base, which may differ from your model’s encoding. Include relevant instructions and context in your planning assumptions; visible response text does not reveal hidden reasoning or other billable output. A sample count is not a complete provider invoice or a prediction of future output.
Price tokens and external services separately
Input tokens represent material sent to a model, including instructions, history and retrieved text. Output tokens represent generated material; billable reasoning can require additional accounting according to the provider. A token is a model unit, not a fixed number of words. Prefer recorded billable usage for the selected model.
Enter separate input and output rates in the same currency and per-million-token units. Multiply a per-thousand rate by 1,000 to express it per million. Rates and billing categories vary by model and service; check current provider documentation.
A tool can incur a service fee and model tokens used to prepare or read its result. Account for each once. Two lookups at hypothetical CAD 0.005 each add CAD 0.01 per execution before the associated tokens. These figures are not a provider quotation.
Choose retry assumptions that fit the process
Use observed expected executions when failures stop at different steps or require different work. The whole-task retry model assumes every attempt has the same expected billable workload and independent, constant failure probability.
At 20% failure and at most three total attempts, expected attempts are 1 + 0.2 + 0.04 = 1.24; final success is 1 − 0.008 = 0.992. One hundred tasks at hypothetical CAD 0.07 per attempt cost CAD 8.68 and yield 99.2 expected successes. Dividing gives CAD 0.0875 per expected success. This does not predict exactly 99.2 completions in one month.
An outage, invalid input or shared model limitation can make successive failures related. In those cases, estimate repeat work from observations and assess success separately.
Separate recurring overhead and omitted expenses
Hosting, monitoring and fixed subscriptions do not disappear when usage is low. Add recurring overhead once per month. Keep development and migration costs in a separate budget, or document an amortization period; neither calculator assumes one.
Review taxes, payment fees, currency conversion, storage, retrieval, human quality checks, image/audio processing and support. The token calculator covers text-token charges. The workflow calculator adds only external charges and overhead you enter. Neither reconstructs every invoice line.
Reproduce low, base and high scenarios
Each hypothetical CAD scenario uses one step, 10,000 input tokens per execution, input rate CAD 1 per million and output rate CAD 4 per million. Each execution also makes two external calls at CAD 0.005 each. Fixed monthly overhead is CAD 25.
Executions already include retries: choose Enter expected executions mode. Success assumptions are 95%, 90% and 80%. The high case combines more tasks, repeated work and longer outputs; it is a planning scenario, not a confidence interval.
| Assumption / result | Low | Base | High |
|---|---|---|---|
| Initiated tasks | 800 | 1,000 | 1,500 |
| Executions per task | 1.0 | 1.2 | 1.5 |
| Output tokens per execution | 1,000 | 2,000 | 4,000 |
| Model cost per execution | CAD 0.014 | CAD 0.018 | CAD 0.026 |
| Tool cost per execution | CAD 0.01 | CAD 0.01 | CAD 0.01 |
| Variable monthly cost | CAD 19.20 | CAD 33.60 | CAD 81.00 |
| Fixed overhead | CAD 25.00 | CAD 25.00 | CAD 25.00 |
| Monthly estimate | CAD 44.20 | CAD 58.60 | CAD 106.00 |
| Expected success | 95% | 90% | 80% |
| Expected successful tasks | 760 | 900 | 1,200 |
| Cost per success (rounded) | CAD 0.058158 | CAD 0.065111 | CAD 0.088333 |
Verify the arithmetic and interpret unit costs
Base model cost is (10,000 × 1 + 2,000 × 4) ÷ 1,000,000 = CAD 0.018. Tools add 2 × 0.005 = CAD 0.01, giving CAD 0.028 per execution. Then 1,000 × 1.2 × 0.028 + 25 = CAD 58.60. Expected successes are 1,000 × 0.9 = 900; 58.60 ÷ 900 = CAD 0.065111… per expected success.
Low: 800 × 1 × (0.014 + 0.01) + 25 = 44.20. High: 1,500 × 1.5 × (0.026 + 0.01) + 25 = 106.00. Costs per initiated task are CAD 0.05525, 0.0586 and 0.070666… respectively.
Successful-task cost allocates all spending across expected outcomes, including failed work. If expected successes are zero, the ratio is undefined. At zero initiated tasks, overhead can still produce a positive total.
Total = N × Σ(ej × [(Ij × Pin,j + Oj × Pout,j) / 1,000,000 + Σ(calls × rate)]) + F; expected successes = N × sUse the calculators in sequence
Start with the AI Token Cost & Monthly Budget Calculator for one request or a text-token budget. Its affordability calculation finds whole requests that fit a token budget; external fees and project overhead are excluded.
Use the AI Workflow & Agent Cost Calculator for multiple steps. Enter the base case, open External charges within the step and add two calls at 0.005. Keep CAD and explicit executions selected. Download the result to retain the entered assumptions.
Visit the AI Cost Intelligence hub to choose the scope. Compare separate scenarios using consistent success definitions.
Replace assumptions with billing evidence
Reconcile pilot estimates with usage records and the invoice period. Check whether cached input, long-context tiers, batch processing, tool fees or minimum charges use different rates. Record effective dates and contractual terms alongside your inputs.
Average tokens multiplied by average executions can mislead when long tasks retry more often. Prefer total billable tokens divided by total executions for each step, or split materially different task types into separate estimates. Expected costs do not describe the range or timing of actual spending.
Keep an assumptions log: task definition, sample period, volume, per-step tokens and executions, rate source and currency, tool fees, success criterion, overhead and exclusions. Recheck it when the workflow changes. Provider references below were checked October 17, 2026 for billing categories; no live rate is embedded.