Five disjoint token categories
Token counts measure text and other information processed by a model. For coding agents, context includes instructions, conversation history, source snippets and tool results as well as the final response. The total is not simply the number of words written for you.
Mr. Usage separates uncached input, output, cache reads, five-minute cache writes and one-hour cache writes. Each token belongs in only one category. For OpenAI records, cached input is subtracted from total input before calculating uncached input; reasoning tokens already included in output are not added a second time.
How an API estimate is calculated
For each known model, multiply each token category by its published price per million tokens, divide by one million, and sum the results. Cache-hit prices can be much lower than new-input prices, while output is often more expensive. Two workloads with the same total tokens can therefore have very different API-equivalent values.
For example, at input/output prices of $2/$10 per million and cache reads of $0.20, one million uncached input tokens, 100,000 output tokens and five million cached tokens cost an estimated $4: $2 input + $1 output + $1 cache reads. The 6.1 million-token total is not priced as if every token were output or new input.
Subscription value versus API spend
API-equivalent value answers: what would this recorded workload cost at the selected model’s API rates? It does not answer how much your subscription charged you, whether a plan guarantees that amount of work, or how much money you saved.
The website separates its Subscription API-value board from the API spend board. API-spend values are estimated token charges for usage categorized as API, not verified invoices. Historical local logs cannot reliably prove a billing category. A person may appear in more than one category when distinct usage buckets are shared.
The current native app shares supported usage together on the website’s API-value board. Older or self-hosted integrations may expose category settings. Neither website ranking verifies spending, identity or productivity.
What the estimate excludes
The website uses a versioned model-price catalog and standard short-context rates, including cache pricing. It excludes tools, fast-mode premiums, long-context surcharges, regional multipliers, discounts and taxes. The native app may account for additional per-response details where recorded, so its estimate can differ from the website’s.
Unknown models still contribute to token rankings but are excluded from dollar totals and identified as unpriced. OpenAI All devices category/model splits are estimates. Provider prices can change, and previously uploaded snapshots retain their computed prices until a later sync. Treat these figures as monitoring estimates, not financial records.
Based on the app’s published documentation. Provider behavior and features can vary by installed version. Read as Markdown.