Savings tracker

AI coding cost & savings tracker

Pick your tool, set how your work splits across task difficulty, and see your annual cost reduction — broken down by task complexity, by model, and by supported tool.

Most bills come from one habit: reaching for the top-tier model on every call. Cheaper triages each request and routes it to the cheapest model that still clears the bar — capped by whatever model you asked for, so it only ever downgrades or matches, never spends more. Hard, correctness-critical work (concurrency, security, proofs) is protected from being downgraded, not escalated past what you requested. Estimate that below, then see exactly where the money goes.

Estimate your savings

Your task mix — how your requests split by difficulty
Estimated annual savings
$0
Baseline assumes you run the top-tier model on every request. Prices are illustrative output rates ($/Mtok) with equal token weight per task; your real numbers come from the gateway monitor. Simple work drops to a model ~5–20× cheaper; complex work stays on the frontier.

Where your money goes

Same estimate, three views of the cost reduction. Use the projection slider to scale it from a single month out to two years — it defaults to a full year.

1 moProjection: 12 months (1 year)24 mo

Test it on your real history

The estimate above is a model. Want your actual numbers? After installing, run one local, read-only command — cheaper peek scans your existing .claude, .codex, … chat logs and reports exactly what routing would have saved, in tokens and dollars. Nothing leaves your machine; it classifies prompt text only, never your files or tool output.

npx cheaper peek                 # scan all history
cheaper peek --days 30              # just the last 30 days
cheaper peek --harness claude-code  # one tool  ·  add --json to script it
cheaper peek — what adaptive routing would have saved
Example output — not your data
scanned last 30d across your harness chat logs

harnesscallsdowngradabletokensyou’d save
● Claude Code1,284876 (68%)18.4M$41.20
◐ Codex412223 (54%)5.1M$7.80
○ CursorDB-backed (not yet readable)

Total  1,696 calls · 1,099 downgradable · from you 1,204 / sub-agents 492
Spent on record   $71.40
Could have saved  $49.00 (69% off) · 14.9M tokens re-routable
Annualized        $596/yr (extrapolated from 30d)
Example output — not your data
Biggest opportunities (top-tier calls that didn’t need it):
  $0.38  opus→haiku   you   rename getUser to fetchUser across the repo
  $0.31  opus→haiku   sub-agent   list the files in src/components

Example output shown above — not your data. The real command reads your own logs locally; the web tracker above is the same math without the file access.

How it works with your tool

Install, then start the gateway — it speaks two protocols, so which base URL you set depends on your tool:

npx cheaper install --all
cheaper gateway start
cheaper monitor

Anthropic-native tools (Claude Code, the Anthropic SDKs):

export ANTHROPIC_BASE_URL=http://localhost:8787

OpenAI-compatible tools (Codex, Cursor, Qwen, Gemini's OpenAI-compatibility mode, and most others — see docs for the full list):

export OPENAI_BASE_URL=http://localhost:8787/v1

Read the docs → · See every supported tool →