One real month

This is a different kind of number from the Terminal-Bench floor and the Controlled A/B. It is not a benchmark. It is the author’s own savings ledger over the last 30 days of everyday work, with all six levers running. It shows the shape of the saving when everything fires. It does not predict yours.

Every figure on this page is the same 30-day window (2026-08-15 to 2026-09-14, snapshot taken 2026-09-14) from the same local ledger, and your dashboard computes the identical report from your own sessions.

What it saved

≈ $1,396

API-equivalent value, 30 days. Three tiers, labelled by how each is known: $126 logged action by action, $1,124 of repeat reads avoided, $146 of other estimates.

59.6M tokens

Never sent to the model. About 28% of the month’s workload prevented: the month cost 150.8M tokens, and the same work without Token Optimizer would have cost about 210.4M.

Tier	Value	How it is known
Logged actions	$126	Metered. Each saving logged as it happened: a result archived, a repeat read skipped, a lean resume, a loop cut short. 2,733 events.
Repeat reads avoided	$1,124	Estimated. Content Token Optimizer removed would have been re-read on every later turn until the session’s next real compaction. 2,153 removals, modelled per turn at that turn’s cache-read rate, capped at the real compaction event.
Other estimates	$146	Estimated. Lean session resumes, compressed sub-agent context, MCP output caps, loops prevented, hint-served reads. Each listed below.
Total	$1,396	Tokens are the primary figure; dollars depend on which models you run and are shown second for that reason.

Metered means each saving was logged as it happened. Estimated means the saving is modelled from what happened. Estimates are always shown separately and never dressed up as metered.

The same month in tokens

	Tokens	Share
Spent (fresh input plus de-duplicated output)	150.8M	71.7%
Saved, metered	25.4M	12.1%
Saved, estimated	34.2M	16.2%
Would have been sent without Token Optimizer	210.4M

The same 30 days, in full

Sessions

Measure	Value
Sessions	1,698 main sessions, plus 104 sub-agent sidechains
Platform	Claude Code CLI (the author’s primary platform)
Average prompt-cache hit rate	67.2%
Quality-scored sessions, all-time	5,210 (ledger begins 2026-05-30)

Logged actions, per lever (metered, $126)

The directly-metered tier: output Token Optimizer shrank, evicted, or never let through, logged with before and after token counts as the sessions actually ran. Dollars are repriced to each session’s own model mix.

Mechanism	Events	Tokens removed	Saved
Tool-output archive	1,113	18.80M	$100.09
Archive re-fetch blocked	103	2.38M	$12.71
Lean-output nudge (measured)	291	1.21M	$6.42
Bash output compression (pipelines, git, generic)	883	2.73M	$5.83
Loop detection	124	0.12M	$0.35
Checkpoint restores	113	0.08M	$0.34
Cross-turn dedup	23	0.11M	$0.21
Structure maps (re-reads)	4	9.5K	$0.05
Delta reads	5	−0.5K	−$0.00
Quality nudges	74	0	$0.00
Total	2,733	25.43M	$126.00

Delta reads ran at a small net loss this month: five events where the diff was longer than the file. It is reported as measured, not hidden.

Lean-output nudge. When context fills past 25%, a short cache-safe note (injected as additionalContext, never rewriting the existing prefix) tells the model to reason deeply but keep its visible output lean. Each nudge records the session’s pre-nudge average output tokens, then meters the actual post-nudge output against it: max(0, pre-nudge average − post-nudge average) × post-nudge turns. Never an estimate, never negative, and never booked until post-nudge turns exist.

Behavioral savers. Quality nudges and loop detection prevent waste instead of removing it. Both log an event when they fire and are low-volume by design; their payoff is avoided degradation and avoided retry loops, not a headline dollar figure.

Repeat reads avoided (estimated, $1,124)

Measure	30 days
Removals counted	2,153
Tokens removed at the time	20.5M
Re-read tokens avoided on later turns	1.99B
Value at each turn’s cache-read rate	$1,124.13

Only metered-magnitude removals count here. Every removal is multiplied by the de-duplicated number of turns until that session’s next real compaction event, priced at the per-turn cache-read rate, and stops at the compaction. Lean resumes and other estimated-magnitude events are excluded so nothing is counted twice.

Other estimates ($146)

Source	Volume	Tokens	Saved
Lean session resumes (cold reload avoided)	156 resumes	31.85M	$105.16
Sub-agent runtime compression	846 dispatches	5.60M	$27.99
MCP output cap	466 capped results	2.33M	$11.72
Loops prevented (continuations not run)	131 events	0.13M	$0.67
Hint-served reads	2 hints used	10K	$0.05
Total			$145.60

Quality grades

The 7-signal quality score rolls up to an S/A/B/C/D/F grade. Across the 1,698 main sessions in this window:

Grade	Sessions	Reading
S	714	Exceptional: minimal waste, high decision density
A	210	Good: clean context, efficient tool use
B	463	Normal: some bloat, recoverable
C	77	Degraded: significant waste, coaching recommended
D	162	Poor: heavy bloat, likely retries or loops
F	72	Failing: near-total waste

The grade scale is identical on every platform, so grades compare across hosts even when the underlying signal count differs (3 to 7 signals).

First-read skeleton promotion

Large code files on first read that are unlikely to be edited soon can be served as a skeleton, with the full original archived and retrievable via expand, a ranged Read, or a direct Edit. A cohort (language plus size band) is only promoted to active serving after proof from real history: the edit-within-5-turns rate must stay under 15% across 20 or more reads in 5 or more distinct sessions.

A live tripwire watches every active cohort. If a cohort’s live edit-after-skeleton rate ever crosses 15%, it auto-demotes to measure-only and logs a cohort_demoted event. Demotions are sticky to prevent flapping; re-promotion is explicit (measure.py cohorts promote <lang:band>) or happens on the next history backfill. The full original is always archived before any skeleton is served, and if archiving fails the full file is served unchanged (fail-open).

As of this snapshot the tripwire has demoted all four code cohorts: live edit-after-skeleton rates of 16.0% (Python 16-64KB), 16.7% (Python 64-256KB), 50.0% (TypeScript 16-64KB) and 40.0% (TypeScript 64-256KB), each above the 15% gate. The history backfill still rates the same cohorts promotion-ready on their all-time edit rate (0.6% to 1.5%), which is exactly the disagreement the tripwire exists to catch: live behaviour wins. Markdown cohorts stay measure-only by policy, because a headings-only outline drops load-bearing prose. The first-read retarget that shipped in v2 serves the file you opened in full and moves the compression budget to skeletons of the related files around it; see Read cache. Both flags are visible and toggleable in measure.py v5 status.

What is excluded

The benchmark harness writes its own test-fixture events into the same local ledger. In this window that was 766 events (7.2M tokens, $14.47) tagged f3bench-session-0001 and test-session-*, recorded between 2026-08-31 and 2026-09-06. They are excluded from every figure on this page.

Reproduce it on your own data

# The savings ledger behind the tiers above (last 30 days)
python3 scripts/measure.py savings
python3 scripts/measure.py savings --json

# Live compression events from your history database
python3 scripts/measure.py compression-stats --days 30 --json

# Cohort status and tripwire state
python3 scripts/measure.py cohorts status --json

# Full dashboard (the same report, visualized)
python3 scripts/measure.py dashboard

Your number is your own. Savings scale with how much read-heavy, multi-file work you do. Someone who mostly chats will see less. Someone who lives in large codebases with long sessions will see more. Token Optimizer computes this same report from your own history, so you never have to take this month’s figure on trust.