Record the conditions beside the number
Use the exact release, operating system and instrument. Record the folder’s trust state, the account type and the model. Each can change what the session sends.
← All insights
KN-08 · Token measurement standard
When can you trust a token number?
A token count describes a particular setup. In two Claude Code sessions, the usage panel and the log were close on cache reads but disagreed on output by about 16 and 23 times.
In this article
Six conditions for a repeatable measurement
- Release. A harness update can change the startup payload between sessions.
- Operating system. In the Windows test, PowerShell added an estimated 3,369 tokens and came enabled by default.
- Instrument. The stripped-down terminal started at 1,884 tokens and the desktop app at 8,633. Treat cloud as a separate setup too.
- Folder trust. The setup reported that an untrusted folder withheld allow rules while retaining deny rules, hooks and environment settings. Record the trust state and test behavior at your release.
- Account type. Two accounts can receive different prompts at the same release.
- Model. In one folder, the startup payload for one model contained about 6,000 more characters than for another. This comparison was in characters.
The study used one Windows machine. Repeat a measurement on a second machine and operating system before calling it reproduced across systems. Compare the same tool or file block across setups before comparing totals.
Two records of the same calls
The desktop app displayed session usage. The harness log recorded the usage returned by the API for each call. The team compared the two instruments across two sessions.
| Session and measure | Panel | Log |
|---|---|---|
| Opus 5.5 · cache read | 152.8M | 152,565,400 |
| Opus 5.5 · output | 28.4k | 461,837 |
| Fable 5.1 · cache read | 5.3M | 5,253,398 |
| Fable 5.1 · output | 3.6k | 84,821 |
Figure 01Field measurement
The instruments agreed much more closely on cache reads
Output tokens
The log reported about 16× and 23× the rounded panel counts. Both model charts use a 0–500,000 output-token scale.
Cache reads
The rounded panel and log counts were much closer. Both model charts use a 0–160,000,000 cache-read-token scale; the scale changes with the measure.
Each model represents one session on the same Windows machine. Panel numbers are rounded. Output counts differed by about 16× and 23×. The study did not establish the cause or compare charges.
Check the counting rule first
- Deduplication. The log sum used one value per message. For the Fable session, choosing the first, last, largest or smallest record per message always gave 84,821.
- Plausibility. The Fable session visibly produced about 150,000 characters across text and tool input, compared with 3,600 output tokens on the panel.
- Speed. Using the panel’s elapsed time, the log totals implied about 70 output tokens per second for both sessions. The panel totals implied about three or four.
Counting only the last call per turn and counting text alone did not explain the difference. The cause remains unknown.
Use the right source for billing
Anthropic describes the Claude Code session cost figure as a local API estimate. API customers should check the Claude Console for authoritative billing. Subscription usage and billing follow the applicable plan. This comparison tested telemetry and did not evaluate charges.
Publish enough detail to check the number
- Record the six conditions and the date.
- Name the instrument and the unit of deduplication.
- Cross-check discrepant totals against a source record. Report an unknown cause plainly.
- For output analysis, use a deduplicated call log and recheck it against the current release.
How we know
Two sessions, two models, one day in late September 2026, and one Windows machine. The panel gap was not rerun on the current release or for this publication. Resumed or forked sessions can copy history into a new file, and side threads can keep separate records. Linux and macOS were outside the study.
Related: What you pay before your first word.