A per-token price list looks simple until you try to predict what one useful task will cost. The headline rate is only part of the calculation: you also need the billing unit, the input and output split, and the number of requests your workflow actually makes. Take this first lap through a price list before funding a larger run.
1. Find the billing unit before reading the rate
Start with the column label. A price might be quoted per thousand tokens or per million tokens. Those numbers are not directly comparable until you put them on the same scale. Also check the currency rather than assuming every table uses dollars.
Write down three things: the model name, the currency, and the token unit. If a row says a rate applies per million tokens, divide your token count by 1,000,000 before multiplying by that rate. Do not multiply the full token count directly by the displayed price.
2. Separate input tokens from output tokens
Input tokens are the content sent to the model. Depending on the tool, that can include instructions, conversation history, source files, retrieved documents, and tool results. Output tokens are the content the model generates. A short question can therefore sit inside a much larger billable request.
Record the input and output rates separately. If you are checking Claude Sonnet 5.5 or Gemini 2.5 Flash, use the exact matching row rather than borrowing a rate from another model in the same family. Similar names do not establish identical billing.
3. Read the conditions attached to each column
Before calculating, look for qualifications such as context-length tiers, cached input rates, or separate charges for particular modalities. These are possibilities to check, not features you should assume every price list offers.
- Context tiers: Does the rate change when the request crosses a stated threshold?
- Caching: Are cache writes and cache reads listed separately, and what qualifies for each?
- Additional usage: Are images, audio, tools, or other features priced outside the text-token columns?
If the table does not explain a condition, check its billing documentation before treating the lowest visible number as your expected rate.
4. Calculate one request with explicit assumptions
Use this formula for ordinary text input and output: request cost = (input tokens / billing unit × input rate) + (output tokens / billing unit × output rate). Add any separately documented charges afterward.
Here is a fictional example, not a SwiftRelay quote. Assume input costs $0.80 per million tokens and output costs $3.20 per million tokens. A request containing 12,000 input tokens and generating 1,500 output tokens costs $0.0096 for input plus $0.0048 for output: $0.0144 total. At exactly that usage, 1,000 requests would cost $14.40.
5. Count the requests behind one completed task
A completed coding task may involve several model calls, not one. An assistant can inspect files, call a tool, receive its result, and request another response. Later calls may include earlier conversation content again.
Budget for the whole task. Measure a small sample of representative runs, then record total input tokens, total output tokens, and total cost per completed task. Include retries and unsuccessful attempts. Multiplying only the first request by your expected task volume can understate spending.
6. Check measured usage before expanding the run
SwiftRelay is a practical option when you want pay-as-you-go access across supported models with per-token billing. Sign up for SwiftRelay and use the small gift credit provided to new accounts for a limited, representative test.
Compare your estimate with the usage records available through your tools and account. Investigate differences in token counts, cache treatment, or additional charges before increasing volume. Keep the model, rate, billing unit, and test date together in your notes: a useful cost estimate is a reproducible calculation, not just a copied headline price.