A person in a black cap faces a round vault door lined with blue and violet glass cylinders, where an orange boxy robot stands at a podium.

Field note. Version 1.23.0

The Cache-Read Tax

Across 344 AI agent sessions and 45,876 turns measured on Sep, 26 2026, cache reads accounted for 78.1% of cost-weighted token spend and generated output for 10.6%, against an average of 341,348 tokens of context re-read on every single turn.

Collected to . Sample: 344 sessions, 45,876 turns, one organization.

Key findings.

1. Across 344 AI agent sessions and 45,876 turns measured on Sep, 26 2026, cache reads accounted for 78.1% of cost-weighted token spend.
2. In the same 45,876 turns measured on Sep, 26 2026, generated output was 0.27% of tokens by raw count and 10.6% of cost-weighted spend.
3. Across 344 agent sessions at one organization measured on Sep, 26 2026, each turn re-read an average of 341,348 tokens of context.
4. Cache writes accounted for 11.3% of cost-weighted token spend across 344 agent sessions measured on Sep, 26 2026.

The data.

  • Generated output, raw token count0.27%
  • Generated output, cost-weighted10.6%
  • Cache write, cost-weighted11.3%
  • Cache read, cost-weighted78.1%
The same 45,876 turns measured two ways. Generated output is 0.27% of tokens by raw count and 10.6% of spend once the classes are weighted by price; cache read dominates either way.
Measured on the Orbyt Labs repository, Sep, 26 2026. Each figure is produced by the command in its source column.
MetricValueHow it is counted
Sessions measured344node packages/collective/src/token-efficiency.mjs --json
Turns measured45,876same run
Average context re-read per turn341,348 tokenssame run, avgCtx
Cost-weighted spend: cache read78.1%output 5x, cache-write 1.25x, cache-read 0.1x vs base input
Cost-weighted spend: cache write11.3%same weighting
Cost-weighted spend: generated output10.6%same weighting
Generated output by RAW token count0.27%unweighted counts, the misleading figure
Worst single session, cost-weighted burn170,298,489same run, session c653d1f1

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 0.27% of tokens, which reads as noise and suggests that shortening what an agent writes is pointless. Cost-weighted, it is 10.6%, 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 10.6% figure.
  • Cost-weighted share is not the same as opportunity. Output being 10.6% 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. This table records version numbers and changes.

VersionDateChange
1.23.0Refresh: sessions 326 -> 344; turns 45,000 -> 45,876; avgCtx 344,261 tokens -> 341,348 tokens; cacheReadShare 78.2% -> 78.1%; cacheWriteShare 11.2% -> 11.3%.
1.22.0Refresh: sessions 284 -> 326; turns 42,841 -> 45,000; avgCtx 350,884 tokens -> 344,261 tokens; cacheReadShare 78.5% -> 78.2%; cacheWriteShare 10.9% -> 11.2%.
1.21.0Refresh: sessions 239 -> 284; turns 40,570 -> 42,841; avgCtx 356,821 tokens -> 350,884 tokens; cacheReadShare 79.0% -> 78.5%; cacheWriteShare 10.6% -> 10.9%; outputShare 10.4% -> 10.6%; outputRawShare 0.26% -> 0.27%.
1.20.0Refresh: sessions 217 -> 239; turns 39,681 -> 40,570; avgCtx 363,329 tokens -> 356,821 tokens; cacheReadShare 79.4% -> 79.0%; cacheWriteShare 10.4% -> 10.6%; outputShare 10.2% -> 10.4%; outputRawShare 0.25% -> 0.26%.
1.19.0Refresh: sessions 213 -> 217; turns 39,502 -> 39,681; avgCtx 364,102 tokens -> 363,329 tokens; cacheReadShare 79.5% -> 79.4%.
1.18.0Refresh: sessions 209 -> 213; turns 39,302 -> 39,502; avgCtx 364,932 tokens -> 364,102 tokens; cacheWriteShare 10.3% -> 10.4%.
1.17.0Refresh: sessions 204 -> 209; turns 39,101 -> 39,302; avgCtx 365,860 tokens -> 364,932 tokens; cacheReadShare 79.6% -> 79.5%; outputShare 10.1% -> 10.2%.
1.16.0Refresh: sessions 203 -> 204; turns 39,061 -> 39,101; avgCtx 366,046 tokens -> 365,860 tokens.
1.15.0Refresh: sessions 200 -> 203; turns 38,934 -> 39,061; avgCtx 366,637 tokens -> 366,046 tokens; cacheWriteShare 10.2% -> 10.3%.
1.14.0Refresh: sessions 193 -> 200; turns 38,639 -> 38,934; avgCtx 368,003 tokens -> 366,637 tokens; cacheReadShare 79.7% -> 79.6%; worstSession 122,931,945 -> 170,298,489.
1.13.0Refresh: sessions 191 -> 193; turns 38,567 -> 38,639; avgCtx 368,355 tokens -> 368,003 tokens.
1.12.0Refresh: sessions 190 -> 191; turns 38,522 -> 38,567; avgCtx 368,572 tokens -> 368,355 tokens; cacheReadShare 79.8% -> 79.7%; cacheWriteShare 10.1% -> 10.2%.
1.11.0Refresh: sessions 182 -> 190; turns 74,020 -> 38,522; avgCtx 364,839 tokens -> 368,572 tokens; cacheReadShare 75.9% -> 79.8%; cacheWriteShare 11.7% -> 10.1%; outputShare 12.4% -> 10.1%; outputRawShare 0.32% -> 0.25%; worstSession 279,332,673 -> 122,931,945.
1.10.0Refresh: sessions 175 -> 182; turns 72,860 -> 74,020; avgCtx 366,027 tokens -> 364,839 tokens; cacheWriteShare 11.6% -> 11.7%; outputShare 12.5% -> 12.4%.
1.9.0Refresh: sessions 199 -> 175; turns 91,077 -> 72,860; avgCtx 373,126 tokens -> 366,027 tokens; cacheReadShare 75.3% -> 75.9%; cacheWriteShare 12.1% -> 11.6%; outputShare 12.7% -> 12.5%; outputRawShare 0.33% -> 0.32%.
1.8.0Refresh: turns 91,076 -> 91,077; avgCtx 373,127 tokens -> 373,126 tokens.
1.7.0Refresh: sessions 169 -> 199; turns 81,672 -> 91,076; avgCtx 376,364 tokens -> 373,127 tokens; outputShare 12.6% -> 12.7%.
1.6.0Refresh: sessions 166 -> 169; turns 80,824 -> 81,672; avgCtx 377,387 tokens -> 376,364 tokens.
1.5.0Refresh: 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%.
1.4.0Refresh: sessions 158 -> 164; turns 78,598 -> 79,996; avgCtx 380,013 tokens -> 377,920 tokens; cacheReadShare 75.5% -> 75.4%.
1.3.0Refresh: sessions 157 -> 158; turns 78,228 -> 78,598; avgCtx 380,535 tokens -> 380,013 tokens; cacheReadShare 75.4% -> 75.5%.
1.2.0Refresh: sessions 153 -> 157; turns 77,165 -> 78,228; avgCtx 382,088 tokens -> 380,535 tokens.
1.1.0Refresh: 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.
1.0.0First publication.

More from Research.

The other measurements from the same repository, and the papers they sit beside.

Cite this.

Bartak, J. (2026). The Cache-Read Tax. Orbyt Labs Research, version 1.23.0. https://www.orbytlabs.ai/research/agent-token-economics

CC BY 4.0. Reuse it with attribution.

Back to all research