Claude Code and Cursor Costs: How to Budget AI Coding Agents Before You Pull an Uber

Quick answer

Budget AI coding agents per developer, not per seat. Anthropic’s documentation puts the enterprise average for Claude Code at around $13 per developer per active day, or $150-250 per developer per month, and Cursor Teams starts at $40 per user per month before on-demand usage. The averages are manageable. What breaks budgets is the tail: a minority of heavy agentic users, overflow billing switched on by default, and no per-developer cap. Uber found that out the expensive way.

This post covers what Claude Code and Cursor actually cost in October 2026, what happened at Uber, where the bill hides, and a five-step budget model that holds. All plan prices below were checked on vendor pricing pages and docs on 9 October 2026.

What happened at Uber?

Uber is the case everyone in engineering leadership is quoting, so it is worth getting the facts right.

• Uber’s CTO, Praveen Neppalli Naga, told The Information that the company had used up its full-year 2026 AI budget within the first few months of the year, and was “back to the drawing board” on its spending assumptions. Coverage attributes the overrun mainly to rapid Claude Code adoption, with Cursor usage reported as plateauing (AI Magazine, Benzinga).
• Engineers had been encouraged to use the tools, and coverage reports internal leaderboards that ranked usage.
• Uber later confirmed to Bloomberg a $1,500 monthly cap on employee token spending for agentic coding tools such as Claude Code and Cursor (Outlook Business).

Various outlets have published per-engineer cost figures for Uber, but they disagree with each other and none comes from Uber directly, so we have left them out.

The lesson is not “AI coding tools are too expensive”. It is that Uber measured adoption and rewarded usage before it measured cost per developer and capped the tail. That is a FinOps sequencing failure, and it is the same mistake teams made with cloud a decade ago.

What does Claude Code cost per developer?

There are two ways to pay, and they behave very differently.

Sources: Claude pricing, Claude Code cost docs.

The number to anchor on is Anthropic’s own: “Across enterprise deployments, the average cost is around $13 per developer per active day and $150-250 per developer per month, with costs remaining below $30 per active day for 90% of users.” Read that last clause carefully. Ten percent of users sit above $30 per active day, and those are the people who decide whether you hit budget.

What does Cursor cost for a team?

Sources: Cursor pricing, Cursor Teams pricing docs.

Two details in the Teams docs matter for budgeting. First, on-demand usage is enabled by default on Teams and billed in arrears, with third-party models charged at public API list prices plus a Cursor token rate. Second, admins can set a monthly team-wide spending limit, but per-member spend limits are only available on Enterprise. On Teams, one heavy user can eat the whole team’s headroom.

Where does the AI coding bill actually hide?

The seat price is the visible part. These are the drivers that move the invoice:

• Overflow billing. Cursor on-demand is on by default; Claude usage credits let members keep working past their allowance at API rates. Both are useful. Both remove the natural ceiling a seat plan gives you.
• Model defaults. Anthropic’s docs point to Opus left as the default model and sessions that are never cleared as the usual causes of unexpectedly high API spend. Defaulting to a cheaper model for routine work is the single easiest saving.
• Long sessions and cache misses. Claude Code resends the conversation with each request. A session left open all day, or resumed after the cache expires, reprocesses a large context. Our LLM API cost optimization guide covers the caching maths behind this.
• Multi-agent runs. Anthropic documents that agent teams use roughly 7x more tokens than standard sessions when teammates run in plan mode. Useful for big tasks, expensive as a habit.
• Attribution gaps. If developers reach Claude through Bedrock or Vertex, Anthropic’s analytics do not see that usage, and your cloud bill shows it without a developer attached. You need a gateway or OpenTelemetry to know who spent what.

How should you budget AI coding agents?

Here is a worked budget for a 100-developer team, using Anthropic’s documented averages and a deliberately explicit tail:

The cap figure is a policy choice, not a vendor number. The point is that the heavy segment gets its own line and its own ceiling, instead of hiding inside an average. Then run the five steps:

1. Baseline a pilot cohort. Ten to twenty developers across seniority levels for four weeks. Record cost per developer per active day, by model. Anthropic’s docs recommend exactly this before wider rollout.
2. Pick the billing model per segment. Standard seats for most, premium seats or metered usage for a named heavy group, and an explicit decision on whether overflow is on.
3. Cap at the narrowest level available. Claude Team and Enterprise support spend limits at organisation, group and member level. Cursor Teams only supports a team-wide cap, so if per-developer control matters, that is an argument for Enterprise or a gateway.
4. Alert on velocity, not month-end totals. A looping agent can burn a week’s budget in an afternoon. Daily per-developer alerts catch it; a monthly report does not. This is what our budget and alert validation work tests.
5. Show cost back to teams. Publish each team’s number monthly. Start with showback rather than chargeback; our showback vs chargeback guide explains why.

Which spend controls exist for each setup?

Sources: Claude Code cost docs, Cursor Teams pricing docs.

If your developers use more than one of these paths, and most organisations do, you need one place that reconciles them. That reconciliation is the core of an AI cost audit, and it is where the tools in our AI and GPU cost tools ranking differ most.

Is the spend worth it?

That is the right question, and it is the one Uber’s leadership was publicly asking. Uber’s CEO said on the Q1 2026 earnings call that about 10% of the company’s code is now written and submitted by AI agents, as reported by Outlook Business. Share of code is an activity metric, though, not a value metric. Before your next renewal, pair spend per team with something that matters to the business: cycle time, escaped defects or tickets closed. If cost rises and none of those move, the leaderboard is measuring the wrong thing.

Where to start on Monday

Turn off default overflow for everyone except a named heavy-user group, set a cheaper default model for routine work, and switch on daily per-developer spend alerts. Those three changes take an afternoon and remove most of the Uber-style risk. Then run the four-week baseline before you expand access.

If you want an outside view, our AI coding-agent spend audit, delivered through our AI and GPU cost governance service, maps every path developers use to reach Claude, Cursor and other agents, attributes cost per developer and team, and tests that your caps and alerts actually fire. Talk to us before the budget conversation happens without you.