Costs & cache

The Costs page is for understanding where your token budget went and what to do about it. Three sections, all driven by the same SQLite the Sessions and Heatmap pages read.

Headline

A single big number — total cost across the selected range — with a week-over-week delta.

$84.21          ↘ -12% from last week
This week · 47 sessions · 3,128 prompts · 71% cache read rate

Color-flagged: green if down, red if up. A small sparkline shows the trend across the range.

Per-prompt cost ranking

A sortable list of the top 50 most expensive prompts in the range. Each row shows:

Column	Meaning
Order	Rank in the list.
Prompt text	Truncated to ~120 chars.
Cost	This prompt’s cost_usd (sortable column).
Tokens	Input / output / cache split (micro-bar).
Model	The model that handled it.
Cache hit %	This prompt alone.
Session	Clickable link to the Sessions deep-dive scrolled to this prompt.

Click any row to expand inline — full prompt text, every tool call with its result-token cost, the token-breakdown table.

The pattern you’re looking for isn’t always “one prompt that cost $5” — sometimes it’s “five prompts of $1 each, all asking variations of the same thing because the agent didn’t have enough context the first time”.

Per-day line chart

A simple time series across the range (default 30 days), one point per day, hover for daily total + prompt count.

The line tells you when you spend. Spikes correlate with launch crunches, big refactors, the “explored a new codebase” day. The flat baseline tells you what your steady-state cost looks like.

By-model breakdown

A donut chart showing total spend share between models:

• Opus 4.6 / 4.7 / 4.8
• Sonnet 4.5 / 4.6
• Haiku 4.5

For most teams, the donut tells the most actionable story:

Pattern	Action
80%+ Opus	Are you using Opus for read/edit-only turns? Route those to Sonnet — same outcome, ~5× cheaper.
50/50 Opus/Sonnet	Healthy split for most workflows.
80%+ Sonnet	Either you’ve optimized well, or you’re not using Opus for the right things (planning, hard refactors).
Significant Haiku spend	Usually a hidden subagent — Haiku is the default for fast lookups. Check if it’s appropriate.

Cache analytics

A dedicated card with one big number — your cache read rate for the range — plus a 2×2 breakdown:

Cell	What
Creation tokens	Total tokens written into the cache. Split into 1-hour and 5-minute lanes (Claude’s two cache TTLs).
Read tokens	Total tokens served from cache. Charged at ~10% of normal input rate.
Saved this week	Estimated dollars saved vs no cache.
WoW delta	Cache rate change versus last week.

The cache read rate is the single most controllable knob for keeping Claude Code cheap. Three concrete plays:

Play	Effect
Pin a stable system-prompt prefix	The first N tokens of every conversation hash the same → those tokens come from cache forever after.
Use CLAUDE.md for shared context	Same idea at the project level — every session in this project sees the same memory, so the cache prefix is huge.
Avoid “tell me about” prompts	Each one writes a unique prefix to cache that may never be read again. Use structural queries (specship_explore) that return reusable answers instead.

A team running Claude Code 20+ hours a week typically sees the cache read rate move from ~40% to ~75% after adopting these — about a 2.5× reduction in real input-token cost.

The Costs page lives in the desktop UI — start it with specship serve --ui and open Costs. It shows the per-day spend, the week-over-week change, a by-model breakdown, and the top prompts by cost. The Compare page breaks each project’s cost down per model and lists its top tools.

Compare projects

A sibling page — Compare projects — takes the cost data and pivots it across projects. Each row is a project; columns are total cost, sessions, avg cost/session, cache hit rate, top 3 tools used, drift queue size.

The “most efficient project” callout highlights the project with the best cache_hit_rate and lowest avg_cost_per_session — a target for the others to match.

This is the surface team leads use for weekly cost reviews.

→ Next: Tips engine — pattern-matching over your transcripts.