What it costs to keep a Claude Code project running.

Plan for two recurring lines, not one. The first is the Anthropic-API line: across enterprise deployments Anthropic publishes an average of $13 per active developer per day and $150 to $250 per developer per month, with 90 percent of users staying under $30 per active day. That is the meter on Claude Code itself. The second line is the engineering retainer: the work the meter does not cover, including CLAUDE.md and skills upkeep, model migrations, escalation paths for the 20 percent of cases the agent should not handle alone, and reconciliation when other writers (humans, formatters, cron jobs) touch the repo. On c0nsl.com the published retainer band is $1,000 to $5,000 per month. Verified against Anthropic's cost docs and the c0nsl rate card on 2026-05-04.

The published per-developer figure assumes a developer typing into Claude Code on an active day. An SMB project in maintenance mode has a different shape: maybe one engineer touches it twice a week, the agent runs a few scheduled jobs, and a teammate or formatter occasionally rewrites bytes. Most months the API spend will land lower than $150 because there are fewer active days. But the architecture-tax on CLAUDE.md, skills, and the four-files contract still has to be paid, even on weeks where nobody opens the IDE. That is why the maintenance bill is more usefully expressed as 'API spend plus retainer' rather than 'X per developer per active day.'

Anthropic's official guidance is to keep CLAUDE.md under 200 lines because it loads into context at every session start. At roughly 80 tokens per line, a 200-line CLAUDE.md is about 16,000 input tokens per session. Without prompt caching, on Sonnet input pricing that is about $0.05 per session of CLAUDE.md alone; over 30 sessions per month, around $1.50. With prompt caching active and the prefix re-read instead of re-written, reads are billed at 0.1x base, which drops the same 30 sessions to a fraction of that. The dollar number is small. The point is not the size of the bill, the point is that CLAUDE.md is a recurring cost that compounds with skills, MCP tool definitions, and any extra detail the file accumulates over the life of the project. Trim it the way you would trim a hot loop.

Anthropic's own cost docs say agent teams use approximately 7x more tokens than standard sessions when teammates run in plan mode, because each teammate maintains its own context window and runs as a separate Claude instance. If your maintenance workflow spawns three or four parallel teammates to handle a refactor, the API line for that day can swing from $13 to closer to $90 even before extra runtime. This is the single biggest variance source between a calm month and a noisy one. If you do not need fan-out, do not enable it; agent teams are off by default and require setting CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1.

Inside any session, run /usage. It prints total cost, total API duration, total wall duration, and lines added or removed. The dollar figure is a local estimate computed from token counts and may differ from your authoritative bill, which lives at platform.claude.com/usage in the Console. For team-wide spend, set workspace rate limits on the Limits page in the Console; that is also where you cap Claude Code's share of your overall API budget so a runaway session cannot starve other production workloads. On Bedrock, Vertex, and Foundry, Claude Code does not send metrics back, so several large enterprises have used LiteLLM as an unofficial spend tracker.

Yes, but it is small. Anthropic notes that even idle sessions consume tokens for background functionality such as conversation summarization for the --resume feature and command processing. The published number is typically under $0.04 per session. That is not a budget item, it is a footnote. The real recurring spend lives in the active sessions, the skills the agent loads, the MCP servers you connect, and how often the codebase is re-read.

The parts no token meter measures. CLAUDE.md, skills, and the four-files contract drift over the life of a project: writers change, the codebase grows, the model gets a new release, the MCP server list shifts. Someone has to read the drift and rewrite the architecture. The retainer also covers the 20 percent of cases that should not be automated, the human-escalation surface for emotional or compliance-sensitive workflows. And it covers vendor-side maintenance, including OAuth-scope reviews after the 2026 Vercel and Context.ai breaches, prompt-caching tuning when input shapes change, and the model-migration scaffolding that keeps Opus 4.7 calls from breaking when the next version ships. None of this lives on the API meter, all of it has to be paid by someone.

Significantly, on workloads that have any repetition. Cache writes with the default 5-minute TTL are billed at 1.25x base input, cache reads are billed at 0.1x base. The break-even is two requests in the same window, not two thousand. For a maintenance workflow where the same CLAUDE.md, skills, and file headers are loaded on most sessions, caching turns a recurring 1x bill into a recurring 0.1x bill on the cached portion. That is also where the most common silent bug shows up: a templated value in a system prompt (a per-user variable, a current timestamp) that invalidates the prefix on every call. If your monthly bill is bigger than expected, the first thing to check is whether the cache is actually firing.

A worked example. One engineer touches the project two days per week (about 8 active days per month), Sonnet for most work and Opus for occasional architecture passes, prompt caching active on the system prompt and skills. At Anthropic's published $13 per active day average, that lands around $104 per month on the API meter. Add roughly 25 percent variance for the occasional agent-team fan-out, prompt-cache misses on a templated value, and one heavy debugging session, and a realistic monthly API line is $130 to $160. Add a $1,500 per month engineering retainer for everything the meter does not cover. Total maintenance: roughly $1,630 to $1,660 per month for a maintained SMB project.

Launch is concentrated, maintenance is diffuse. During launch the engineer is in a session most working hours, agent teams might fan out for parallel work, prompt-caching savings have not yet stabilized because the prompts themselves are still being shaped, and the cost shape looks closer to the upper end of the per-developer band. After ship, the active-day count drops, the system prompt and skills stabilize, prompt caching settles into its 0.1x read groove, and the meter falls. What does not fall is the architecture-tax on CLAUDE.md, the writer-drift reconciliation when other entities touch the repo, and the 20 percent of cases that route to a human. Those become the dominant cost.

Trim CLAUDE.md and move workflow-specific instructions out of it into skills, which load on demand instead of on every session start. Anthropic's own guidance puts the ceiling at 200 lines for exactly this reason. Most CLAUDE.md files I read on existing client repos are double that size, with hundred-line sections describing PR-review or migration workflows that are needed once a month. Moving those into a skill saves real tokens on the 95 percent of sessions that do not invoke the workflow. Combined with prompt caching on what remains, this single change can cut the API line by 20 to 40 percent on a mature project.