Skip to main content

AI Usage > Input tokens (uncached)

Written by Tom Williams

Dataset: AI Usage

Entity: AI Usage Measurement

Field ID: input_tokens_uncached

Type: Number

Description: The number of input tokens processed without hitting the prompt cache.

Source: App

Transformation logic: N/A

App Mapping

GitHub Copilot

token_usage.prompt_tokens_sum - command line and desktop app records only

Anthropic Claude (coming soon)

tokens.input on Claude Code records; uncached_input_tokens on direct API traffic

Cursor (coming soon)

the sum of tokenUsage.inputTokens across the record's usage events

OpenAI Platform (coming soon)

input_tokens less input_cached_tokens

OpenAI Codex (coming soon)

N/A - the Codex analytics API publishes no token counts

Reporting Use Cases

The Input tokens (uncached) field is the part of the prompt that had to be processed from scratch. It is the most expensive kind of input token, so it is the one to watch.

  • Consumption trends: Chart it alongside cache reads to see whether growing usage is being absorbed by the cache or is genuinely new work.

  • Per-model comparison: Filter to breakdown = SURFACE_MODEL and group by model_ref to see which models carry the heaviest prompts.

  • Not available for Copilot IDE work: GitHub publishes tokens for the command line and desktop app only. An empty value on IDE records means not reported, not zero.

Did this answer your question?