# 00 — Executive Summary (https://jackin.tailrocks.com/research/token-optimization/00-executive-summary/)



# 00 — Executive Summary [#00--executive-summary]

Every headline number below survived local reproduction, primary-source re-fetch, or
explicit ESTIMATE arithmetic; sources are in each file's Verification ledger.

## TL;DR [#tldr]

* **No honest 10x** exists at zero quality loss today; current defensible stack math lands at
  ≈2.5x (≈2.4x code-heavy), or ≈5–6.2x only if the routing flip passes the harness.
* **Output + cache writes dominate dollars**, even though cache reads dominate token volume; visible
  prose compression is useful but cannot touch thinking tokens.
* **Negative-cost wins are context architecture wins:** tool search, context editing, observation
  masking, Edit-diffs, repo maps, and advisor escalation save tokens while improving quality.
* **Every serious number needs local validation** because session mix, effort level, and tokenizer
  family change the economics materially.

## The verdict on 10x [#the-verdict-on-10x]

**Defensible today: ≈2.5x (≈2.4x on code-heavy mixes). Defensible after validation on your task mix: ≈5–6.2x. A true 10x at
provably equal quality does not exist yet.** The paper path (Sonnet/Haiku main loops with
frontier-model escalation, effort-tiered, context-edited, batch-staged) reaches \~10x in
arithmetic but its quality parity on the hardest tasks is unmeasured (T4) — and the binding
constraints are structural: (1) **thinking bills as output and no style layer touches it** —
only the effort parameter and not-being-the-frontier-model move it; (2) a **cache-read floor**
of context the agent genuinely needs; (3) quality risk concentrates exactly where the remaining
multipliers are. Full math: 30-composed-stacks.md.

## Where the money actually goes (measured, this environment) [#where-the-money-actually-goes-measured-this-environment]

The measured heavy Fable 5 session decomposes as: **cache reads 32% / cache writes 29% /
thinking \~20% / visible output \~17% / uncached input 2%** (02). An independent session
was output-heavier (output 44% / cache writes 34% / cache reads 21%), so use the point estimate
as a profile, not a law. Three consequences the market hasn't priced in:

* **The optimization target is upside down.** Folklore optimizes visible prose (17%); the big
  four-fifths is cache traffic + thinking. One visible-output token = 5 input = **50 cache-read**
  tokens ($50 vs $1/MTok).
* **Thinking is invisible and majority-of-output** (54.8% max-effort main loop; 44.8% across a
  25-agent fleet — local). Claude Code transcripts redact it; it must be inferred as
  `output_tokens − count_tokens(visible)`.
* **Defaults already bank \~4–5x**: caching alone measured −86.3% input-side ($71.59 paid vs
  $524.23 uncached-equivalent, this very session); MCP schemas defer by default; Edit-diffs are
  default. Most "10x easy wins" advice re-sells these defaults.

## The stack (what to actually run) [#the-stack-what-to-actually-run]

**Day 1, riskless (≈1.06–1.3x):** dedup the double-registered caveman hooks (−966 tok/session,
−118/prompt); pin exploration subagents `model: haiku, effort: low` (÷10 on code-heavy text, up to
÷13–14 on prose/markdown-heavy text — re-count the actual corpus); two CLAUDE.md guard-lines (Edit-not-Write; no restatement —
89.3% and 91.4% per-instance, measured); never switch model/effort mid-session (cache is
model-scoped; ≈9-turn break-even per switch).

**With validation (≈2.5x):** effort high→medium on routine work (T1: Opus 4.5 at medium matched
Sonnet 4.5's best SWE-bench with **76% fewer output tokens** — the single strongest sanctioned
number; transfer to Fable 5 must be validated, and this is the only lever that reaches
thinking); context editing / observation masking (vendor: **84% token cut, +29% performance**;
JetBrains T2: masking ≈ −50% cost at parity); register compression on visible prose only
(caveman-ultra measured **58.5%**, not the marketed 65–75%); route half the work
Sonnet-main+advisor (T1: **+2.7pp AND −11.9% cost**; code-heavy routes get list-price ratios,
not the prose tokenizer bonus); batch the offline 30% at 50% off.

**The negative-cost set** (saves tokens AND improves output — adopt unconditionally): tool
search/schema deferral (85% cut, accuracy 49%→74%), context editing, observation masking,
Edit-diffs (aider: quality 20%→61%), advisor escalation, repo-maps/outlines instead of file
dumps (−85/−92% local), effort max→high. Common thread: **less junk in, less junk out** — the
input-architecture layer is where cost and quality align, and it beats every style trick.

**Make it infrastructure (jackin'):** bake the pack into every launched container — env
defaults in the launch env assembly, `[token_policy]` in role manifests, model/effort flags via
CapsuleConfig, CI linter failing when always-loaded context grows. Automatic beats disciplined;
insertion points are mapped in 32.

## Corrections that reorder the field (the graveyard, abridged) [#corrections-that-reorder-the-field-the-graveyard-abridged]

Full kill-tables live in files 10–19; the ones that change decisions:

1. **"Caveman cuts \~75%" → 58.5% measured** (token-level, Fable tokenizer). The 75% (now "65%"
   on the repo) is character-level folklore. And it only touches visible prose — in tool-heavy
   agent sessions, free-text was **1.4–1.5% of visible output** (local); style compression's
   end-to-end ceiling there is \~0.4% of output tokens. In chat-heavy sessions it's real (\~10%
   of dollars). Wenyan: 80.9% char cut collapses to 56.6% tokens — no advantage over ultra,
   higher risk.
2. **"Editing CLAUDE.md mid-session busts the cache" → false**; it's read once at session start.
   Eight real invalidators are enumerated in 13.
3. **"Keepalive pingers save the cache" → solved problem**: Claude Code main loop already
   writes 1h-TTL cache (320/320 calls observed) and "the cache is refreshed for no additional
   cost each time" it's read. Also: count\_tokens does NOT warm the cache (documented).
4. **"1M context costs a premium" → dead**; flat per-token pricing across the window on
   Fable 5/Opus 4.8/Sonnet 4.6 (live pricing page). Quality, not price, is the long-context tax.
5. **"YAML/TOON halve JSON" → minification is most of it**: pretty→minified JSON −29%, →CSV
   −34% further; TOON ≈ CSV+4%; indent width and CSV-vs-TSV are token-identical. Biggest
   structured-data lever is the format spread: pretty XML→CSV = 2.45x.
6. **"RouteLLM saves 85%" → MT-Bench-only** (45% MMLU, 35% GSM8K); per-request gateway routing
   also breaks Claude's model-scoped cache.
7. **"LLMLingua 20x in front of the API" → QUALITY-TRADE trap for coding**: a cache-breaking
   proxy must beat \~5.5x compression to break even vs 0.1x reads; a 2026 RCT on Sonnet 4.5
   found keep-20% compression *increased* cost 1.8%; code tolerates \~10% prompt reduction (T2).
8. **"Compaction is free" → billed as a separate full-price iteration** (\~$1.98 per pass at the
   docs' own example scale), excluded from top-level usage fields.
9. **"Mem0/semantic caches for agents" → evidence is FAQ/consumer workloads**; files-only
   baselines beat Mem0 on LoCoMo; zero coding-agent evidence. Base64/gzip "compression" costs
   2.7–4.3x MORE tokens (measured).
10. **Tokenizer counts are not portable**: Fable 5 = Opus 4.8 tokenizer ≠ Sonnet 4.6 = Haiku 4.5
    (exact family equality, measured); the \~30% premium is an ASCII/English tax, but code/CJK can be
    near-neutral while SCREAMING\_SNAKE is extreme. Cross-model budgets must re-count, and an open docs contradiction
    on prior-turn thinking retention (18 §TL;DR) is worth real money on long sessions.

## What we still don't know (highest-value open measurements) [#what-we-still-dont-know-highest-value-open-measurements]

Quality-vs-effort curve for Fable 5 (the 76% transfer); thinking-share by effort level;
persona-vs-instruction durability over 100+ turns; caveat-drop rates in terse output registers;
composed-stack quality at n=30 (31-validation-harness.md §7 closes all five).

— Read 30 for the math, 32 for the sequence, 31 to prove any of it on your own tasks.