AI token cost: the budget line engineering teams miss

AI token cost is the usage bill a team runs up every time a coding assistant or an agent reads code and writes an answer. Input tokens pay for everything the model is shown; output tokens pay for everything it writes; both are priced per million and charged on top of the people who direct the work. For an engineering team that codes with AI, the cost of the work is no longer just salaries. It is people plus tokens.

That second line is the one budgets miss. A license looks like a fixed monthly fee, so finance plans for seats and stops there. Then agents start running longer tasks, context windows fill with whole repositories, and the usage invoice arrives larger than the seat count suggested. This article explains what drives token spend in a team that codes with AI, how to estimate it before it shows up on an invoice, and the levers that move it. The thesis is simple: tokens are not a reason to avoid AI in engineering, but they are a real cost, and teams that treat them as one keep the gains.

Why AI token spend surprises engineering teams

Token spend surprises teams because it grows with how the tools are used, not with how many people use them. A seat price stays flat; a token bill rises when an agent loops, when more code goes into the context, or when the model writes long answers. The seat count tells finance almost nothing about the usage line.

Four habits make the bill grow faster than expected:

• Output costs more than input. On Anthropic's published API pricing, output tokens cost five times as much as input tokens across the current models; for example, $2 per million input tokens and $10 per million output tokens for Claude Sonnet 5.5, as of October 2026. A model that writes long explanations, full files or repeated drafts spends on the expensive side of the meter.
• Context is paid for every time it is sent. An assistant that reads a large part of the repository before answering pays for those input tokens on each request, unless the platform caches them.
• Agents call the model many times per task. One request from a person can become dozens of model calls as an agent plans, edits, runs tests and retries.
• The same text can cost more tokens on a newer model. Anthropic notes that its newer models use a tokenizer that produces approximately 30% more tokens for the same text. A model upgrade can change the bill even when the work stays the same.

None of these shows up in a seat price. All of them show up in usage.

How AI coding tools charge a team: per seat, per use, or both

Most AI coding tools now combine a seat price with a usage allowance, and charge or block anything beyond it. Some tools are pure usage: you pay per token through an API. The pricing model decides whether token spend is predictable, capped or open-ended.

GitHub Copilot is a useful example because its documentation spells the structure out. According to GitHub's Copilot plans and its usage-based billing documentation, as of October 2026, Copilot Business is $19 per seat per month and includes 1,900 AI credits per user per month, pooled across the organization, with one credit worth $0.01. When the pool runs out, administrators choose whether additional usage is allowed at published rates or blocked until the next billing cycle, and they can set budgets at the user, cost-center, organization and enterprise levels.

API access works differently. Anthropic, for instance, bills per million tokens, with the input and output prices on its model pricing page. There is no seat, so there is no ceiling unless the team sets one.

Cursor, Codex, Claude Code and other tools each publish their own mix of plans and limits, and those plans change often. The rule that holds across all of them: check the vendor's current plan page the day you budget, and write the date next to the number.

How to estimate token spend before the invoice

The reliable way to estimate token spend is to measure it on real work. Pick a handful of representative tasks, run them for a week with the tools your team will actually use, read the usage reports, and multiply by the volume of similar work you expect. Estimates built from seat prices alone will be wrong.

‍

‍

A simple calculation shows how the pieces add up. Suppose an agent session on a mid-sized change reads 200,000 input tokens and writes 20,000 output tokens on a model priced at $2 per million input tokens and $10 per million output tokens. That session costs $0.40 for input plus $0.20 for output: $0.60. Ten such sessions per engineer per day, across a team of six, over twenty working days, comes to $720 a month. Change the model, the context size or the number of retries, and the figure moves fast. The example is arithmetic, not a benchmark; your own pilot week is what gives you the real numbers.

Three things to record during that pilot:

• Tokens per task type, not just the monthly total. Writing tests, reviewing a pull request and refactoring a module have very different profiles.
• Input versus output. If output dominates, the levers are different from a bill driven by context.
• Retries and loops. An agent that repeats calls on a failing test is the single fastest way to burn a budget.

Eight levers that move your team's token bill

Token spend responds to a small set of decisions. Most of them are configuration and team habits, not new tools.

1. Match the model to the task. Anthropic's own guidance is to use smaller models for simple tasks and larger ones for complex reasoning. Not every autocomplete needs the most capable model.
2. Keep the context lean. Send the files the task needs, not the whole repository. Every unnecessary file is paid for on every request.
3. Cache what repeats. On Anthropic's API, reading from the prompt cache costs 10% of the standard input price on most models, so instructions and code that stay the same across requests get much cheaper.
4. Cap agent loops. Set a limit on how many times an agent can retry before a person looks at what is failing.
5. Choose seat or usage on purpose. A seat with an allowance makes spend predictable; pure usage scales with the work. Pick the one that matches how evenly your team uses the tools.
6. Set budgets per team or per person. Budgets turn a surprise invoice into an alert. Tools that support user-level or cost-center budgets make this a setting, not a policy document.
7. Batch what can wait. Anthropic's Batch API, for example, discounts both input and output tokens by 50% for work that doesn't need an immediate answer, such as bulk test generation or documentation passes.
8. Count review time as part of the cost. Cheap code that takes a senior engineer an hour to review is not cheap. The review line belongs in the same budget, and the cost structure of an AI-augmented team lays out how it adds up with people, tools and tokens.

Who should own the token line

The token line needs an owner on the engineering side, not only in finance. Finance can see the invoice; only engineering can see which tasks, tools and habits produced it. Without that owner, the bill gets reviewed once a quarter, after the money is spent, by people who can't change what drove it.

Owning the line doesn't mean approving every request. It means three routines that take minutes a week:

• A monthly look at usage by team and by task type. The usage reports most tools provide are enough to spot the agent that loops or the repository that sends too much context.
• A budget with an alert, not just a cap. A hard cap that blocks the tools mid-sprint creates its own cost. An alert at a set share of the budget gives the team time to adjust.
• Token spend read next to what the work delivered. A rising bill on a team that ships more is a different conversation from a rising bill on a team that ships the same. The number only means something next to the output it paid for.

The owner is also the person who decides when a model upgrade, a new tool or a new agent workflow goes into the budget, because each of those can change the bill without anyone writing a line of code differently.

Frequently asked questions

How much text is 1,000 tokens?

A thousand tokens is about 3,500 characters of English text, based on Anthropic's rule of thumb that one Claude token represents roughly 3.5 English characters. The ratio varies with the language and the tokenizer: Anthropic notes that its newer models produce approximately 30% more tokens for the same text, so the same page can cost more.

How much do 1,000,000 AI tokens cost?

The price of a million tokens depends on the model, and input and output are priced separately. On Anthropic's pricing as of October 2026, Claude Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens. Smaller models cost less, larger ones more, and cached or batched requests get discounts.

Who pays for AI tokens?

The organization that holds the account pays for the tokens, either inside a plan or as metered usage. With a seat plan such as GitHub Copilot Business, each seat includes a monthly allowance of AI credits; with direct API access, every token is billed to the company's account, with no ceiling unless it sets one.

Can AI run out of tokens?

Yes, a team can run out of tokens when a plan's allowance or a budget limit is reached. On GitHub Copilot, for example, administrators decide whether usage beyond the pooled credits is allowed at published rates or blocked until the next billing cycle. A single request can also exceed the model's context window.

‍

Keeping token spend from eating the gain

Token spend eats the gain when it is treated as an afterthought: approved as a license, discovered as an invoice. It stays under control when it is part of the business case from the first day, with an owner, a budget and a usage report someone actually reads.

DevLatam's page on AI-augmented teams puts the condition plainly: “There's a token cost, and it's not zero. Smaller than the headcount it replaces, but a new line in your budget.” The same page frames the model as a squad designed to need fewer people over time, which is exactly why the token line matters: as headcount comes down, tokens become a larger share of what the work costs.

Not every team should take that trade. DevLatam's page on staff augmentation makes the opposite case just as directly: “Token spend is a budget question. Some teams would rather pay a salary they can forecast.” Both answers are legitimate. The mistake is choosing one without pricing the other.

The practical next step is to run the pilot week, put the token line next to the salary line, and decide with both numbers on the table. If you're still deciding which way to staff the work, the three ways to fill the role compare the options, including the one where the team is designed to get smaller.