AI Coding Spend Index
See where your setup falls across four dimensions — cost, caching, context usage, and audit score. For team leads who want a quick read on which dimension is furthest out of line before digging into any one of them.
The math: your inputs are placed against the index's reference bands per dimension. Placement is directional — a prioritization aid, not a market statistic.
Model prices verified 2026-07-02 against provider pricing pages.
| Metric | P25 | P50 | P75 | P90 |
|---|---|---|---|---|
| Cost per dev per day | $4.20 | $8.50 | $18.30 | $42.00 |
| Cache hit ratio | 15% | 38% | 62% | 78% |
| MCP tools per session | 3 | 8 | 14 | 22 |
| Context fill | 25% | 48% | 72% | 88% |
| Audit score | 42/100 | 61/100 | 78/100 | 89/100 |
How does your setup compare?
Run a free audit to see where you stand against these benchmarks and get actionable recommendations.
Run free auditHow to read the results
Read the four dimensions relative to each other, not as absolute grades. A high-spend, low-caching placement points at prompt caching as the first fix; high context usage points at CLAUDE.md and MCP overhead. The band you land in matters less than which dimension is the outlier.
FAQ
- How is my placement computed?
- The values you enter are compared against fixed reference bands per dimension. No account or usage data is read — the placement is a pure function of your inputs.
- Where do the bands come from?
- They are maintained reference bands, not a live survey of teams. Treat placement as directional prioritization, not a claim about where the market sits. For measured data, our analysis of 100 public configs found a median CLAUDE.md of 964 tokens / 92 lines, with 24% of files over 200 lines.
- What are the limitations?
- Your inputs are self-reported estimates, and the bands can't account for workload shape — a research-heavy team and a boilerplate-heavy team can both be efficient at very different spend levels. Verify spend against your provider invoice.
- I landed in a high-spend band — now what?
- Check the concentrated failure modes first: in the 100-config corpus, the top 10 of 100 files held 97% of the detectable waste, and the #1 anti-pattern was session logs inside CLAUDE.md. A short config cleanup usually moves the needle before any model or plan change.
- When should I use this vs the full audit?
- Use this to decide where to look. The full stack audit then measures your actual setup — CLAUDE.md, MCP servers, caching — and produces specific, ranked fixes instead of a band placement.
For your full setup, run the free stack audit — or see the 100-configs report.