Your Biggest AI Cost Isn't Your App. It's Your Team.

Everyone optimizes their application's API calls. But in 2025, 55% of all departmental enterprise AI spend, $4 billion, went to coding tools. Not chatbots. Not AI features in your product. Coding tools: the ones your team uses every day, burning through tokens at a thousand times the rate of your production apps.

The $4 billion nobody planned for

According to Menlo Ventures' 2025 State of Generative AI in the Enterprise report, a survey of nearly 500 U.S. enterprise AI decision-makers, coding captured $4.0 billion of the $7.3 billion spent on departmental AI. That's 55% of all departmental AI spend, the largest category by a wide margin. The next closest, IT, took 10%.

A year earlier, coding AI spend was $550 million. It grew 7.3x in a single year. No other category came close.

Claude Code: half of Anthropic's enterprise revenue

To see how fast this market moves, look at one product. In February 2026, Claude Code passed $2.5 billion in annualized revenue. Its revenue more than doubled in the first weeks of the year, and it accounted for more than half of all enterprise spending on Anthropic products.

Source: Constellation Research, February 2026

Much of that is enterprise teams like yours, running coding agents all day.

The 1,000x token gap

Coding tools cost far more than production apps because of how many tokens each interaction carries.

A production chatbot serving thousands of customers might cost $150–600 a month in API fees. Anthropic's Claude Code documentation has put typical team usage at around $100–200 per developer a month, with wide variance between people. For a 50-person team, that's $5,000–10,000 a month, often an order of magnitude more than the production app they're building.

The cost nobody budgeted for

This isn't a theoretical problem. Enterprises are discovering it on their invoices. CloudZero's State of AI Costs 2025 report found:

• 65% of IT leaders report unexpected charges from consumption-based AI pricing.
• Actual costs frequently exceed estimates by 30–50%.
• 34% of companies don't discover cost overruns until the bill arrives.
• 1 in 5 companies report AI budget variance of around 50%.
• Average monthly AI spend rose from $63K to $85.5K, a 36% increase year on year.

Source: CloudZero, "State of AI Costs 2025"

The FinOps Foundation documented a case where finance estimated AI spend at $3 million a year, but a cross-functional audit found the real number was closer to $6 million: double the budget.

Cursor and Copilot

It's not just Claude Code. Every major coding tool is generating large, growing enterprise bills.

Cursor passed $2 billion in annualized revenue in March 2026, doubling in three months.

Sources: TechCrunch, Bloomberg, March 2026

GitHub Copilot moved to usage-based billing in June 2026. Requests now draw down a pool of AI credits by the tokens they actually use, so a seat is no longer a fixed cost: the heaviest users decide what the whole team pays.

Why this is different from cloud cost management

When cloud costs spiraled a decade ago, the industry built FinOps practices, cost dashboards and governance tools. AI coding costs are on the same path, with one critical difference:

• Cloud costs are visible. Cloud bills break down by service and region. AI coding tool bills show one number.
• Cloud costs are predictable. Provision an instance and you know the hourly rate. AI costs depend on what people type and how the model responds.
• Cloud costs are centrally managed. A platform team controls the infrastructure. AI tools run on every developer's laptop.
• Cloud costs scale with usage you chose. AI costs scale with behavior you can't see: one developer's habits can multiply their token use tenfold.

Adoption is outrunning control

Gartner expects three in four enterprise software engineers to use AI code assistants by 2028. Every one of them will generate token costs, and every one of those costs will need attribution and governance.

The question isn't whether you'll need control over AI coding spend. It's whether you'll have it before the next quarterly invoice arrives.

What this means for you

If you're a CTO, CFO or FinOps leader, here's the uncomfortable truth: your production AI application probably isn't your biggest AI cost. Your team is.

Every developer with Claude Code, Cursor or Copilot makes hundreds of model calls a day, each carrying a full context window, tool definitions and often extended thinking. Your application might make 50 calls a day. One developer's coding agent makes 500. That's where the money goes.

And most organizations still can't break that spend down by person, team, tool or project, or tell whether it's producing work or burning budget on dead ends and runaway agent loops.