For a 10-person engineering team, AI coding tools that cost $200 per developer per month a year ago may now cost $3,000 or more. A seed-stage AI infrastructure company told The Pragmatic Engineer that their per-developer spend went from $200 to $3,000 per month in six months. A healthcare company of 500 people had a single engineer spend $1,400 in one Claude Code session. A finance team set $100-per-user monthly limits and watched developers exhaust them in three to five working days.
The cost increase is structural. Agentic AI workflows consume tokens in ways that flat per-seat subscriptions never did. When a developer runs a multi-hour coding session with tool calls, context refreshes, and retry loops, the bill is measured in dollars per minute. Most engineering budgets from 2024 weren’t sized for that.
What changed in 2025 was the mode, not the tools. Autocomplete suggestions had predictable, bounded costs. Agents do not. When you tell an AI agent to “fix the authentication bug,” it reads files, forms hypotheses, retries on failures, queries for additional context, and may run for hours if no human interrupts it. Each of those steps burns tokens. A developer who switches from autocomplete to agents for their entire workflow can hit 50 times the monthly usage without noticing, because the experience doesn’t feel 50 times more intense. It just feels like the tool is doing more.
The teams holding these costs steady share five specific habits. None of them require switching tools or cutting usage outright.
1. Switch human developers from API billing to per-seat plans
The most common cost mistake is putting developers on API-billed access when flat monthly options exist. Claude Code is $20 per month on the Pro plan and $100 on the Max plan. Cursor is $20 per month. GitHub Copilot Business is $19 per user per month. Any of these options puts a ceiling on what a developer can spend in a month, regardless of how many tokens they consume.
API billing makes sense for automated pipelines: CI/CD integrations, batch processing jobs, code review bots running on pull requests. It makes little sense for a developer running interactive sessions all day. The same agentic session that costs $20 total on Claude Code Max can reach $500 via direct API access if the developer isn’t managing context length. The two billing models solve different problems and they should be applied to different workloads.
The move: migrate human developers to per-seat subscriptions, and keep API access for automation jobs that run at defined times with predictable inputs.
2. Set monthly spending caps before you think you need them
The finance company described in The Pragmatic Engineer’s April 2026 report had 2,000 employees. Their $100 per user monthly limit felt conservative when they set it. Developers were hitting that ceiling within three to five working days anyway. The caps worked, but they forced a conversation about what was actually worth the spend.
That conversation is the real value of a cap. Without a hard monthly limit on API spend, long agentic sessions and routine autocomplete suggestions get treated with the same urgency. A developer doesn’t weigh whether a two-hour refactoring session is the best use of a $200 Claude API budget when there is no visible budget to weigh against.
A reasonable ceiling for a small team is $100 to $200 per developer per month for API spending. That is high enough that a productive developer won’t notice it on a normal day. It is low enough to surface when someone is running overnight sessions on problems that a focused afternoon of manual work would have solved faster.
Anthropic, OpenAI, and Google all support budget alerts in their billing dashboards. Setting one up takes five minutes. The first alert is not there to stop the spend. It is there to start the conversation about what the spend was for.
3. Separate agentic mode from autocomplete use
Most of the token cost in AI coding tools comes from agentic workflows, not autocomplete. An autocomplete suggestion pulls a few hundred tokens. An agent that plans a multi-step refactor, calls tools, reads file trees, and revises its approach can consume tens of thousands of tokens in a single session. The cost gap between the two modes is not marginal.
Teams controlling spend treat agentic mode as a deliberate choice, not a default. For complex problems, agents earn their cost: debugging why a test suite is failing across three files simultaneously, or drafting a database migration where the edge cases aren’t obvious until you read the schema. For routine tasks like writing a single function or generating boilerplate that follows an obvious existing pattern, autocomplete handles it faster and for a fraction of the price.
The discipline: decide which mode you are in before you start. If the task fits in under 10 minutes with autocomplete, use autocomplete. Save the multi-agent session for the problem that would otherwise take a day. A useful heuristic is whether you could write a clear spec for the task in three sentences. If you can, autocomplete can probably handle it. If the spec keeps expanding because the edge cases are unclear, that is an agent problem.
4. Consolidate to one primary tool and cancel the overlap
Developer AI tool stacks accumulated through trial subscriptions and upgrade promotions. A team using GitHub Copilot, Cursor, and Claude Code is paying roughly $59 per developer per month for three tools with substantial capability overlap. Cursor surpassed $2 billion in annualized revenue in early 2026, per TechCrunch, with revenue doubling in the preceding three months. 60% of that revenue came from corporate customers. At some point in the last year, the market consolidated around Cursor as the primary tool.
The argument for running multiple tools is that different tools are better at different tasks. The practical problem is that switching overhead is real. A developer who uses one tool well, knows its keyboard shortcuts, understands its context handling, and has built up prompting habits around its failure modes will consistently outperform a developer who picks tools based on the task. The marginal improvement from “the right tool for this task” rarely exceeds the switching cost.
The quickest audit: ask each developer which tool they open first when starting a task. The one that gets used 80% of the time is worth keeping. The rest are subscriptions supporting a secondary option that rarely gets opened.
5. Replace token leaderboards with shipping metrics
Social pressure to consume more tokens appeared at multiple large organizations in early 2026. Meta’s internal leaderboard, called “Claudeonomics,” tracked token consumption across 85,000 employees, reaching 60.2 trillion tokens in a single month, according to The Pragmatic Engineer. Engineers competed for status titles like “Session Immortal” and “Token Legend.” At Microsoft, one engineer described feeling pressure to burn more tokens “to avoid being seen as using too little AI.”
Shopify took the opposite approach. Farhan Thawar, Shopify’s head of engineering, renamed the team’s AI usage leaderboard to a “usage dashboard” and added circuit breakers to cap runaway sessions. The frame shifted from consumption to outcomes.
Token consumption is not a leading indicator of anything except cost. A developer running 200,000 tokens per day with no shipped code is strictly worse than a developer running 50,000 tokens per day who ships a working feature. A June 2025 longitudinal study of GitHub Copilot adoption across 26,317 commits and 703 repositories found no statistically significant changes in commit activity from Copilot use, despite developers reporting subjective productivity gains. The gap between “feels more productive” and “ships more” is exactly what leaderboards hide.
The metric you reward is the metric you get. Track what shipped.
The thread connecting all five
What these five moves have in common is that they force a connection between AI spending and actual output. Without that connection, token costs fill whatever space is available. With it, developers naturally reserve the expensive mode for expensive problems, and the billing dashboard starts to reflect real work rather than background activity.
None of this requires cutting AI tool usage. The teams spending less are mostly spending the same number of developer hours on AI-assisted work. They have just removed the places where tokens get burned on low-value activity because there was no reason not to burn them.
One thing that helps here: look at what actually shipped in the past month and trace it to the tools that contributed. If a developer used 40,000 tokens on a feature that merged and is running in production, that is a very different spend than 40,000 tokens on a refactor that was abandoned on Friday afternoon. The cost looks identical in a billing dashboard. It is not identical. The teams that understand the difference are the ones where cost and value stay connected.
Pick one of the five changes and put it in place this week. The most common starting point is step four: ask your team which tool they actually use, and cancel the rest. The savings fund the tools worth keeping.
References
| Source | Author / Org | Year | Supports |
|---|---|---|---|
| Token spend breaks budgets: what next? | Orosz, The Pragmatic Engineer | 2026 | $200 to $3,000/dev/month in 6 months; $1,400 single-session spend; $100 cap exhausted in 3-5 days |
| Tokenmaxxing as a weird new trend | Orosz, The Pragmatic Engineer | 2026 | Meta 60.2T tokens/month leaderboard; Shopify Farhan Thawar circuit breakers; social pressure at Microsoft |
| Cursor has reportedly surpassed $2B in annualized revenue | Temkin, TechCrunch | 2026 | Cursor $2B ARR, 60% corporate revenue, market consolidation |
| AI-Powered Pair Programming and Development Productivity | Stray et al., HICSS-59 | 2026 | No statistically significant change in commit activity from Copilot adoption |