Field note - version 1.8.0
The Cache-Read Tax
Across 199 AI agent sessions and 91,077 turns measured on 17 September 2026, cache reads accounted for 75.3% of cost-weighted token spend and generated output for 12.7%, against an average of 373,126 tokens of context re-read on every single turn.
Collected 08-27-2026 to 09-17-2026. Sample: 199 sessions, 91,077 turns, one organization.
The data.
- Generated output, raw token count0.33%
- Generated output, cost-weighted12.7%
- Cache write, cost-weighted12.1%
- Cache read, cost-weighted75.3%
| Metric | Value | How it is counted |
|---|---|---|
| Sessions measured | 199 | node packages/collective/src/token-efficiency.mjs --json |
| Turns measured | 91,077 | same run |
| Average context re-read per turn | 373,126 tokens | same run, avgCtx |
| Cost-weighted spend: cache read | 75.3% | output 5x, cache-write 1.25x, cache-read 0.1x vs base input |
| Cost-weighted spend: cache write | 12.1% | same weighting |
| Cost-weighted spend: generated output | 12.7% | same weighting |
| Generated output by RAW token count | 0.33% | unweighted counts, the misleading figure |
| Worst single session, cost-weighted burn | 279,332,673 | same run, session 37f8c9b6 |
How it was measured.
Every figure is one run of the repository's own measurement tool over its local agent session transcripts, read on the date shown. Nothing is estimated and nothing is carried forward from an earlier measurement.
Cost weighting is applied because the raw counts answer a different question than the one that matters. Generated output is weighted 5x, cache writes 1.25x and cache reads 0.1x against base input, reflecting how those token classes are actually priced.
The headline is the gap between two correct measurements of the same data. By raw count, generated output is 0.33% of tokens, which reads as noise and suggests that shortening what an agent writes is pointless. Cost-weighted, it is 12.7%, which is not noise. Both numbers are true; only one of them describes the bill.
This organization's own always-loaded engineering guidance carried the raw-count framing and drew the conclusion that optimizing output addresses a third of a percent of the problem. That conclusion does not survive weighting, and the measurement tool now says so in its own output.
Context re-read dominates under both measurements, which is the finding that holds either way: an agent turn pays for the whole conversation so far, every time, so cost scales with context size multiplied by turn count rather than with anything the agent produces.
What this does not show.
- One organization's sessions on one tool. Session shapes differ enormously between teams, and a team that clears context aggressively would produce a very different distribution.
- A reading on the date shown. The tool also grades this corpus against thresholds this project set for itself, and a grade is a comparison against thresholds this project set for itself, not an industry benchmark.
- The weighting factors are the ones this project uses. They approximate a real price list but they are not a quotation, and a different weighting moves the 12.7% figure.
- Cost-weighted share is not the same as opportunity. Output being 12.7% of spend does not establish that shortening output is the most effective available saving; it only refutes the claim that it is negligible.
- No causal claim is made. This measures where tokens went, not which practices caused it.
Revisions.
This URL is permanent. When the data is refreshed the version bumps and a row lands here, so a citation made today still resolves to the finding it cited.
| Version | Date | Change |
|---|---|---|
| 1.8.0 | 09-17-2026 | Refresh: turns 91,076 -> 91,077; avgCtx 373,127 tokens -> 373,126 tokens. |
| 1.7.0 | 09-17-2026 | Refresh: sessions 169 -> 199; turns 81,672 -> 91,076; avgCtx 376,364 tokens -> 373,127 tokens; outputShare 12.6% -> 12.7%. |
| 1.6.0 | 09-13-2026 | Refresh: sessions 166 -> 169; turns 80,824 -> 81,672; avgCtx 377,387 tokens -> 376,364 tokens. |
| 1.5.0 | 09-13-2026 | Refresh: sessions 164 -> 166; turns 79,996 -> 80,824; avgCtx 377,920 tokens -> 377,387 tokens; cacheReadShare 75.4% -> 75.3%; cacheWriteShare 12.0% -> 12.1%; outputShare 12.5% -> 12.6%. |
| 1.4.0 | 09-12-2026 | Refresh: sessions 158 -> 164; turns 78,598 -> 79,996; avgCtx 380,013 tokens -> 377,920 tokens; cacheReadShare 75.5% -> 75.4%. |
| 1.3.0 | 09-12-2026 | Refresh: sessions 157 -> 158; turns 78,228 -> 78,598; avgCtx 380,535 tokens -> 380,013 tokens; cacheReadShare 75.4% -> 75.5%. |
| 1.2.0 | 09-12-2026 | Refresh: sessions 153 -> 157; turns 77,165 -> 78,228; avgCtx 382,088 tokens -> 380,535 tokens. |
| 1.1.0 | 09-11-2026 | Refresh: sessions 151 -> 153; turns 75,949 -> 77,165; avgCtx 385,641 tokens -> 382,088 tokens; cacheReadShare 73.6% -> 75.4%; cacheWriteShare 12.2% -> 12.0%; outputShare 12.6% -> 12.5%; outputRawShare 0.35% -> 0.33%; worstSession 3,603,230,006 -> 279,332,673. |
| 1.0.0 | 08-27-2026 | First publication. |
More from Research.
The other measurements from the same repository, and the two papers they sit beside.
Cite this.
Bartak, J. (2026). The Cache-Read Tax. Orbyt Labs Research, version 1.8.0. https://www.orbytlabs.ai/research/agent-token-economics
CC BY 4.0. Reuse it with attribution.