▲ FARTHING

The honest token meter for Claude Code. Count every farthing.

Measure your real bill. Run only what pays. Prove accuracy didn't move — without ever busting the cache.

accuracy-gated token control · part of Nadir · LOWEST VIABLE MODEL · VERIFIED

npx @getnadir/farthing audit      # decompose your real Claude Code bill

getnadir.com · The receipts · Why compression loses · Reproduce

---

Farthing is the cache-safe spend controller for Claude Code & Codex — from the team behind Nadir, the verifier-gated LLM router. Nadir picks the leanest model and verifies the answer; farthing measures the actual bill and cuts it without touching the cached prefix.

This repo is the proof. We benchmarked the token-shaving tools everyone reaches for first, on real Claude Code sessions, priced the actual bill — and put farthing head-to-head. One competitor triples the bill while its own dashboard reports a saving. farthing cuts it and never busts the cache.

---

▮ The receipts

Real multi-turn Claude Code sessions on a fresh expressjs/express clone, Opus, 6 identical prompts per session. 4 arms × 3 reps, interleaved, every turn metered. Full method + caveats: results/arms_4x3.md.

arm	median session	vs baseline	cache_create	separation
baseline	$6.468	—	349k	—
combo (farthing+tokopt)	$3.826	−41%	184k	3/3 ✅
farthing	$4.623	−28%	275k	2/3
token-optimizer	$6.956	+7.5%	366k	0/3 ✗
headroom	—	costs more, 10/10 trials	2.4–8× ❌	—

Read these as directional, not exact. Session cost is brutally noisy — baseline alone swung $6.38 → $10.37 (63%) on identical inputs. What makes the result credible isn't the median, it's that an independent metric reproduces the ranking exactly:

by cost:          combo < farthing < baseline < tokopt
by cache_create:  combo < farthing < baseline < tokopt

PROOF, NOT PROMISES

10/10	2.4–8×	0.4%
headroom cost more in every trial	its cache_create inflation	what a perfect command-output filter is worth on a cache-dominated bill

The finding: the only thing that moves a cache-dominated bill is carrying less context — not compressing it (busts the cache), not filtering side channels (too small), but never putting the redundant thing in context in the first place.

---

Why compression loses on Claude Code

Claude Code's cost model is prompt caching. Three price lanes (Opus, per 1M tok):

lane	price	vs base
cache read (stable prefix)	$1.50	0.1×
fresh input	$15	1×
cache create (rewrite)	$18.75	12× the read lane

A stable prefix is served from cache — nearly free. Compression rewrites the prefix to shrink it, which invalidates the cache from the edit point on, forcing Claude to re-create the whole downstream context at 12× the price.

• headroom deletes ~15–41% of some tokens (its dashboard counts this as saved) but shoves the survivors from the $1.50 lane into the $18.75 lane (its dashboard never counts this). Net: bill up 45–282%.
• farthing never touches the cached prefix. It trims the output lane (terse-output) and blocks byte-identical re-reads (read-ledger) — cache-safe savings — while measuring cache churn instead of causing it. Net: −28% median (n=3, directional), cache_create down — never up.

Same axis — cache_create. headroom manufactures it; farthing measures and avoids it. That is the whole idea.

---

▮ The full study

Every number below is reproducible from REPRODUCE.md.

Agentic matrix — 96 runs, tools ON (results/matrix.md)

4 arms × 8 add/fix/change tasks × N=3 on a real calculator project, accuracy-gated.

arm	median $/task	vs baseline	accuracy	verdict
baseline	$0.311	—	23/24	—
farthing terse	$0.296	−5%	24/24	within noise
caveman	$0.364	+17%	24/24	marginal
headroom	$0.581	+87%	24/24	survives

Only headroom's +87% clears the noise floor here. The ±5% and +17% figures are reported for completeness but are not claims — see results/arms_4x3.md for why session cost needs separation tests, not medians alone.

Real project — monitored express sessions (results/express_real_project.md)

task	baseline	headroom	actual bill	headroom dashboard
trace	$0.947	$1.375	+45%	"saved −3%"
feature	$0.687	$1.087	+58%	"saved −11%"
bugfix	$1.293	$2.578	+99%	"saved −7%"

Faithful re-runs (results/faithful_arms.md)

Objections addressed — real caveman plugin (+17%), real multi-turn + token-mode headroom — both confirmed the result.

Full write-up: docs/FINDINGS.md.

---

The tools, placed

tool	axis it attacks	verdict on Claude Code coding
Nadir	which model runs (route + verify)	−60%, 98% quality — the axis that actually moves the bill
farthing	carry less context; cache-safe layers	−28% median (n=3, directional) — cache-safe by construction, never busts cache
caveman	output tokens	+17% — cuts output (~1% of a coding bill) but adds per-turn injection
token-optimizer	command output	no detectable effect — its filter works (98% on npm test) but that's 0.4% of the bill
headroom	input tokens (compress)	+45% to +282% — compresses, but rewriting the prefix busts the cache

The savings live at model choice (Nadir) and honest measurement + cache discipline (farthing) — never at token-shaving a cache-dominated bill.

---

Reproduce it yourself

See REPRODUCE.md. One command each; every run edits only throwaway /tmp copies and costs real Opus tokens on your own account.

git clone --depth 1 https://github.com/expressjs/express.git /tmp/express-bench
( cd /tmp/express-bench && npm install )
headroom proxy --port 8787 --mode token &
bash benchmark/express_session.sh baseline feature
bash benchmark/express_session.sh headroom feature   # compare totals, not its dashboard

---

Layout

benchmark/        the harnesses (run_matrix, express_session, multiturn, run_bench)
bench-calculator/ the calculator target repo (8 tasks, accuracy gate)
results/          computed tables + raw CSVs (synthetic tasks only)
docs/             FINDINGS + method
index.html        getnadir-styled landing page (GitHub Pages)

Method & honesty

• Compare the session total (sum of per-turn total_cost_usd), never a tool's own dashboard — the express test exists because those diverge.
• Cost is noisy (5-min cache TTL, agent nondeterminism): N≥3, median, interleave arms.
• Compare cost only among runs that passed the accuracy gate.
• The single-shot Q&A A/B was run on a private corpus; raw prompts are not included in this public repo. Only aggregate numbers remain.

Built alongside Nadir — the verifier-gated LLM router. route lower · verify · escalate only if needed