Claude Code for teams: what it costs and how to control it
API billing scales with every developer on your team. Here is what that looks like at 5, 10, and 20 seats, and what you can do about it.
The $2,400/month problem
A team of 5 developers running 8 Claude Code sessions per day, across 20 working days, generates 160 sessions per developer per month. At an average of 50,000 input tokens per session, that is 8 million input tokens per developer per month. At $3 per million tokens, each developer costs $480/month. Five developers: $2,400/month.
Per-seat costs at scale
| Team size | Monthly API cost (raw) | After 40% reduction (standard session) | You save |
|---|---|---|---|
| 3 devs | €1,440/mo | €864/mo | €576/mo |
| 5 devs | €2,400/mo | €1,440/mo | €960/mo |
| 10 devs | €4,800/mo | €2,880/mo | €1,920/mo |
| 20 devs | €9,600/mo | €5,760/mo | €3,840/mo |
Savings = raw token cost minus reduced token cost. Based on Claude Sonnet $3/M input pricing. See full pricing.
Why teams spend more than solo developers
Solo developers naturally scope their sessions. A team of developers working in parallel generates simultaneous long running sessions. One developer deep in a refactor, another debugging a production issue, a third building a new feature, all accumulating context independently. The cost scales with headcount, not just usage.
There is also the consistency problem. Solo developers develop intuitions about when to start a fresh context window. Teams do not naturally share those intuitions. Longer sessions become the default.
Per-seat pricing, not per request
The Distillery runs locally on each developer's machine. There is no per request pricing, no data leaving your network before optimization, and no change to the models or responses your team receives. Each developer sets one environment variable. The proxy intercepts, distils, and forwards transparently.
Real world cost reduction is 30-60% depending on session patterns — the calculation above uses the 40% measured standard-session rate; heavier sessions save more (up to ~85%). The exact number depends on how your team uses Claude Code.