Not $20 a Month: What AI Coding Really Costs per Engineer (Uber Caps It at $1,500)

The $20 plan is the floor, not the bill

Ask what AI coding costs and you'll usually hear a plan price. $20 for Claude Pro. $20 for ChatGPT Plus. $10 for GitHub Copilot Pro. Those are real prices, and they are the smallest number you will ever pay. Here is what the cheapest and the top individual plans cost today, checked on each official pricing page on November 10, 2026.

Monthly billing shown; annual billing is cheaper on some plans. Each plan comes with a usage allowance, and the allowance is where the money moves. When an engineer runs out, the choices are to wait for the limit to reset, buy a bigger plan, or turn on usage credits billed at API rates. Teams that pick the last option are the ones whose bills surprise them.

Seat prices, usage terms and the official source for every plan are on the AI coding plans comparison. If you only need one tool, there are pages for Claude Code plans, Codex plans, Cursor plans and GitHub Copilot plans.

What one developer actually spends, by Anthropic's numbers

The best public number comes from Anthropic itself. Its Claude Code cost documentation says that across enterprise deployments the average is about $13 per developer per active day and $150 to $250 per developer per month, and that 90% of users stay below $30 per active day. That is token spend at API rates, not a plan price.

Put that into a year and the $20 plan looks silly. $150 to $250 a month is $1,800 to $3,000 per developer per year. A heavy user who hits $30 on each of 20 working days spends $600 that month, and that is still inside what 90% of users stay under. The other 10% go higher, and nothing in the average tells you how far.

The heavy-user row is my arithmetic, not an Anthropic figure. It shows where the top of the normal range sits, not a ceiling.

What Uber, Apple, Tesla and Accenture allow per engineer

A few companies have now put a number on it, mostly because the bills got big enough to make the news. These are the caps that have been reported, with how solid each one is.

Two things stand out. First, the caps are far above Anthropic's average, which means they're aimed at the heavy users, not the typical ones. Second, the direction is down. Apple cut its cap by a third, and 9to5Mac reports that some teams could go to $100 a day. Uber's cap is per tool, which Simon Willison worked out can still add up to $36,000 a year for one engineer on two tools.

The odd one out is Accenture. Its $29 is what the company pays centrally; anything above goes on a project or team budget. Employees quoted by 404 Media said they spent $500 to $1,000 a month, so the real figure is set project by project. Meta, for now, has no cap at all, though Adam Mosseri told TechCrunch caps will probably be needed in a year or two.

Where the bill blows up

Nobody gets a $4,000 month from typing slowly. Anthropic's own documentation names the usual causes of unexpectedly high spend, and they are boring:

• Long sessions that are never cleared. Claude Code resends the whole conversation with every request, so a one-line question in a session that has been open all day pays for the whole day's history.
• The most expensive model left as the default. A switch to Opus also applies to subagents that inherit the session's model.
• Agent teams. Anthropic says they use about 7 times more tokens than a normal session when teammates run in plan mode.
• Scheduled loops and background tasks, which keep sending the full context even while nobody is at the keyboard.
• Coming back after a break longer than the cache lifetime, which reprocesses the full context at full price.

The pricing itself is also moving under people's feet. OpenAI moved Codex to a token-based rate card in April 2026, and from June 1 GitHub Copilot code review started using GitHub Actions minutes. Neither change shows up on a plan's headline price.

The prompt cache is the quiet one. Our cache hit rate study priced one agent task on 72 models: the median cost fell from $1.05 with no cache to $0.29 at a 90% hit rate. Coding agents resend a lot of context, so the same arithmetic applies to them.

Subscription or API rates: which is cheaper for one engineer

Here's the part that saves money. If an engineer would spend $150 to $250 a month at API rates, a $100 Max plan or a $200 Ultra plan is cheaper, as long as its usage limits hold that engineer's work. If they spend $15 of tokens a month, every plan costs them more than paying per token.

Plan limits are measured in usage, not dollars, so this table is a guide, not a guarantee. The way to know is to run /usage in Claude Code for a week. On a paid plan it shows which skills, subagents and MCP servers use the most, and suggests fixes for anything behind 10% or more of your usage.

For a side-by-side of Claude Code, Codex and Cursor across several usage profiles, with the subscription break-even worked out, see Claude Code vs Codex vs Cursor cost. To price your own workload in tokens, the Claude Code cost calculator does it in a minute.

A budget you can defend

If you have to put a number in a spreadsheet tomorrow, start from Anthropic's range and multiply. It's the only public average from real deployments that we found, and it lands well below every reported cap except Accenture's.

• Run a pilot first. Anthropic's advice is to start with a small group and measure before rolling out, and it's right: your codebase and habits move the number more than any table.
• Set a per-person limit, not only a company total. On Claude Team and Enterprise, admins can set spend limits for the organization, a group or a single member. On the API, workspace limits do the same job for the whole workspace.
• Give heavy users a higher cap on purpose instead of letting them run into a wall mid-task. Your top spenders may well be the people shipping the most.
• Check the bill monthly against the pilot. If spend per engineer is climbing, look at session length and model choice before you cut anyone's limit.

If you'd rather plan in tasks than in seats, the AI budget planner turns a target number of tasks into a monthly spend. And if a hard stop is what you need, how the spending limits of each AI API work covers what actually stops the bill and what only sends an email.

FAQ