Token Efficiency
Where the tokens go, where they're wasted, and exactly how to save them
Act now — policy levers
Admin · one config change · 2.1B tokens/mo recoverable — set it here and the system complies instantly
- Route trivial tasks off OpusModel right-sizingPOLICY
340 short, high-acceptance sessions ran on Opus where Haiku/Sonnet would suffice — Opus bills 5× the quota for identical output.
340 sessionsEffort: lowConfidence: high2.1B tokens/mo
Coach to recover — behavioral levers
Per-engineer · via Craft · 3.5B tokens/mo recoverable — quantified here, but there's no switch to flip: each is fixed by coaching how people work, in Craft
- Trim bloated contextContext bloatCOACHING
120 sessions exceeded 120k tokens of context without a /clear — most of it unused but re-billed every turn.
120 sessionsEffort: lowConfidence: high1.4B tokens/mo→ Craft → Token efficiency - Raise prompt-cache hit rateCache efficiencyCOACHING
Cache hit is 61% vs a 75% target: restarting sessions and re-pasting files throws away the warm cache, so context gets re-billed at full rate.
org-wideEffort: mediumConfidence: medium975M tokens/mo→ Craft → Token efficiency - Reduce abandoned sessionsAbandoned workCOACHING
210 sessions burned tokens but ended with no accepted diff — usually unscoped work that should have started in plan mode.
210 sessionsEffort: mediumConfidence: high700M tokens/mo→ Craft → Plan-mode discipline - Cut repeated failed tool-call loopsRetry loopsCOACHING
Sessions retried the same failing command 5+ times before changing approach, re-billing the whole context each loop.
90 sessionsEffort: lowConfidence: medium425M tokens/mo→ Craft → Iteration efficiency
Realized savings — audited
Projected vs what actually landed after you acted. Realized deltas are MEASURED (before/after token series keyed to each acted move); projections are PROXY counterfactuals. 4 moves still inside their verification window.
- Sonnet downshift · Checkout routing policyRealizedJun 28Projected725M tokensRealized778M tokens(107%)
Tracking 107% of projection — verification window closes Jul 12
- Thrash coaching · GrowthRealizedJun 21Projected230M tokensRealized185M tokens(80%)
Verified Jul 3
- Cache config fix · SRE CorePartialJun 14Projected525M tokensRealized425M tokens(81%)
Variance: 3 engineers switched models manually before rollout
- Onboarding nudge · 11 idle seatsNo effectJun 09Projected578M tokensRealized0(0%)
NO EFFECT: 9 of 11 seats still idle — re-minted with escalation
- Context→CLAUDE.md playbook · DevExRealizedJun 02Projected160M tokensRealized173M tokens(108%)
Verified Jun 18
The one move that saved nothing is shown too — that's the point. Attribution is causal inference, so no-effect rows stay visible with their cause notes rather than being quietly dropped.
Plan fit · sizing signal, not waste
Unused allowance can't be recovered — you've already paid for it. This informs the next renewal.
On plan — tracking to 86% of allowance; 5.2B tokens headroom. This is a sizing signal for renewal, not recoverable waste.
Efficiency signals
Diagnostic vital-signs — how waste is detected. The recoverable tokens and the fixes live in the lever sections above.
Tokens by model
Daily · last 30 days — model mix is the biggest lever, and it feeds the policy lever above
Usage anomalies
Token spikes vs 30-day baseline
- Analytics+62%
Large data-dump prompts spiked Opus usage
Jun 28 – Jul 1+600M tokens - SRE Core team+41%
Long unscoped sessions, no /clear
Jun 26 – Jul 2+58M tokens - search-svc repo+33%
Repeated full-repo reads, low cache hit
Jul 1 – 2+48M tokens
- zero-rejection window14 sessionsSRE Core · team-level
- edits with no prior read9 sessionsVendor Pod Chennai · team-level
- token spike, flat tool diversity6 sessionsAnalytics · team-level
- abandoned-restart chain5 sessionsGrowth · team-level
Conversation starters — faking work costs nearly as much as doing it. Full context in Guard → Alerts.
Token breakdown
Group token usage and optimization headroom by any dimension
| Repo / project | Tokens / mo | Share | Optimizable |
|---|---|---|---|
| web-storefront | 5.3B tokens | 583M tokens(11%) | |
| payments-svc | 4.6B tokens | 1.0B tokens(22%) | |
| ml-pipeline | 4.2B tokens | 1.1B tokens(27%) | |
| mobile-app | 3.1B tokens | 469M tokens(15%) | |
| analytics-etl | 3.0B tokens | 744M tokens(25%) | |
| api-graph | 2.6B tokens | 312M tokens(12%) | |
| auth-gateway | 2.4B tokens | 218M tokens(9%) | |
| search-svc | 2.1B tokens | 409M tokens(19%) | |
| billing-core | 2.1B tokens | 409M tokens(19%) | |
| infra-terraform | 1.8B tokens | 256M tokens(14%) |
Recover by team — vs each team's own baseline
Not a leaderboard. Each team is compared only to its own 90-day baseline, ordered by recoverable tokens — a work queue, top = biggest opportunity to coach.