Skip to content

Token estimates

Anton Aliev edited this page Oct 24, 2026 · 2 revisions

Token estimates · evidence before a number

Русская версия · Home

Necessary request context is the baseline. Whole-file sizes are not a universal substitute.

---
config:
  theme: base
  themeVariables:
    primaryColor: "#CCE4F7"
    primaryTextColor: "#1A1A1A"
    primaryBorderColor: "#2D7FD3"
    lineColor: "#0067C0"
    secondaryColor: "#FAFAFA"
    tertiaryColor: "#F3F3F3"
---
flowchart TD
  task["Alternative task context known?"] -->|No| unknown["Cannot estimate"]
  task -->|Yes| seen["Already in cloud context?"]
  seen -->|Yes| zero["No source-reading credit"]
  seen -->|No| credit["Credit only declared necessary context"]
  zero --> cost["Subtract complete response and request overhead"]
  credit --> cost
  cost --> result["Conditional saving or additional tokens"]
Loading

For an explicit text scenario: S - V - 2R, where S is necessary unseen source context, V the complete visible response and R the request. Local generation is not automatically credited as avoided cloud generation. Unknown provenance or an unknown baseline remains unknown.

For log triage with a custom question, the comparison baseline is UNKNOWN; the entire log is not credited as avoided context. In compiler-log triage, the model-generated answer to that question is reported separately from the strictly verified diagnostic inventory.

Example: a 100,000-token file, a 4,000-token task

If the task needs 4,000 source tokens, the returned response is 600 and the request is 100, the illustrative comparison is 4,000 - 600 - 200 = 3,200, not 99,200. If the source was already read, its credit becomes zero and the same call adds 800 tokens. This example is not a measured result or an account charge.

Try the illustrative calculator · Full methodology and limits.


Source: dedd223a9af7. Changes via pull request.

Clone this wiki locally