A founder asked me last month how much it would cost to build their app with AI. I gave them the answer everyone wants — a range. They paused, then said the number that everyone says: "that's less than I expected." We built the prototype. The actual API and tool bill came in roughly four times my range. Not because anyone did anything wrong. Because the question they asked, and the question I answered, were both about the price of the tools rather than the cost of the work.

This is the conversation I now have with every founder who asks about AI coding economics. The tools are cheaper than building software the old way. The bill is also often higher than the demos suggest. Both things are true. Understanding why requires looking at how AI coding tools actually charge, where token usage hides, and what separates the teams that spend efficiently from the teams that hemorrhage tokens without realizing it.

How AI Coding Tools Charge in 2026

There are three pricing models, and most teams pay through at least two of them simultaneously.

Subscription-based seats. A flat monthly fee per developer for access to a coding assistant — Cursor, Copilot, Claude Code, the various IDE plugins. This covers a baseline of usage. Power users hit the rate limits and either pay overage fees or fall back to a different tier. The headline price is small per seat; the total for a team adds up.

Pay-per-token API usage. When you call the underlying AI model directly — via an agent framework, a custom tool, a CI integration — you pay per input and output token. Modern coding workflows are long-context: an agent reviewing a pull request might process a million tokens for a single review. At even modest per-token prices, this stacks.

Hybrid bundled plans. Increasingly, the dominant model. A subscription that includes a generous monthly token allowance, with overage charges past the limit. The headline cost looks predictable. The actual cost depends on how aggressively the team uses agents and long-context features.

The misleading part is that the per-call cost of any given AI action is small. A code completion costs fractions of a cent. A file review costs a few cents. An agentic refactor across a codebase costs a dollar or two. Every individual call feels free. The bill at the end of the month is in the thousands for active teams. The math is "small times many."

Where Tokens Actually Go

This is the part that surprises everyone, including teams that have been using AI tools for two years.

Context, not output. Most of the tokens in a modern AI coding session are input context — the files the AI is reading to understand what to do — not the output it produces. A simple "add a function to this file" might involve the AI reading thirty files to understand the project, costing twenty times more in input tokens than output tokens. Teams that don't realize this look at their tools and conclude they aren't generating much code, while the bill says otherwise.

Long-running agents. When you tell an agent "fix this bug" and walk away, the agent might run for an hour, reading and re-reading files, trying approaches, backing out, trying again. Every iteration costs tokens. The agent's autonomy is the feature; the autonomy's cost is its appetite. A single complex agentic task can cost more than a day of subscription-tier usage.

Repetitive context. Without good caching, the AI re-reads the same files over and over throughout the day. Modern tools cache context aggressively — but only when configured to. Teams that haven't enabled or properly configured prompt caching are paying full price for the same context many times an hour.

Failed attempts. Every time the AI tries a change, the user rejects it, and the AI tries again — every iteration consumed tokens. A user who is precise about their requests gets results in two iterations; a user who is vague gets results in five. The vague user paid more, generated more frustration, and got worse results.

What Different Workflows Actually Cost

Solo founder building a prototype. A typical AI-built prototype — a basic web or mobile app, maybe a few thousand lines of generated code — runs from a few hundred dollars to a couple thousand dollars in tool costs over four to six weeks. That sounds like a lot until you compare it to the alternative of hiring a developer for the same work, which would be ten to thirty times more. The AI option wins on cost. It does not win on the "free" framing some demos imply.

Small team building a product. A team of two to five developers using AI tools heavily typically spends a few hundred to a couple thousand dollars per developer per month across subscriptions and API usage. The unit economics make sense — the per-developer cost is a small fraction of the developer's salary — but it is meaningfully more than the "$20/month per seat" pricing page would suggest.

Heavy agentic workflows. Teams that run AI agents in CI, use long-running coding agents, or do AI-driven code reviews on every pull request can spend ten times more than the subscription baseline. The value can be there — these workflows replace significant human effort — but the bill is real.

What to Actually Do About It

Track your token usage from the start. Most teams discover their token costs at the end of the first month when the bill arrives. Set up usage dashboards, configure spending alerts, look at where the tokens are actually going. Treating this as observable instead of mysterious is the single biggest cost-management win.

Enable prompt caching everywhere it is available. Modern AI APIs cache common context fragments — system prompts, project context, large reference files — at a fraction of the per-call cost. If your tool supports caching and it is off or misconfigured, you are paying full price for context the system is repeatedly seeing. This single setting can cut bills by 50 to 80 percent for context-heavy workflows.

Match the model to the task. The frontier models are not always the right choice. A code completion task that finishes in a second on a smaller, cheaper model does not need the largest, most expensive model. Tools that let you route different tasks to different model tiers — fast cheap models for autocomplete, capable expensive models for hard reasoning — produce dramatic savings without losing quality.

Write precise requests. Vague prompts produce vague output that needs to be iterated. Each iteration costs tokens. The discipline of describing what you want clearly the first time isn't just better for results — it is meaningfully cheaper. The most expensive prompt is one that produces useless output and forces a retry.

Audit your agentic workflows quarterly. Long-running agents that ran on every pull request when the project had five contributors might cost ten times more when the project has fifty. Workflows that made sense at one scale can become expensive at another. The audit doesn't have to be deep — just enough to notice when a workflow that used to cost ten dollars now costs a thousand.

The Stakes for People Who Care About Margins

For solo developers and small teams, token costs are a real line item but rarely the deciding factor — the productivity gain dominates the cost. The teams that get burned by tokens are the ones that adopt the tools without watching the bill, then suddenly discover a four-figure invoice they did not predict.

For larger teams, token economics start to look more like cloud bills did in 2015 — something you have to actively manage or it gets away from you. The pattern of "developers adopt new AI tools faster than finance can track" is repeating the cloud-cost pattern, and the result will be the same. Eventually a FinOps function emerges for AI spend. The teams that get ahead of this aren't going to look much different on the outside; they will just have predictable margins instead of mystery bills.

The conversation about AI coding economics shouldn't be "this is cheap" or "this is expensive." It should be "what does this cost in our specific workflow, are we paying for value we use, and where are tokens leaking." The teams asking those questions are quietly building cost-aware AI development practices. The teams not asking them are funding their tools through surprise.

The cost is real. The value is also real. Knowing the actual numbers is what separates the teams that benefit from AI development from the teams that just talk about it.