The flat monthly fee for AI coding tools is on its way out. What you pay is shifting to pay-per-use, where the bill rises with how hard you push the tool. GitHub Copilot makes the switch on June 1, 2026. Cursor rewrote its prices back in June 2025. Anthropic has spent the past year tightening Claude Code, its tool that writes and edits code for you. The reason is the same everywhere. One click from a user can now cost the company wildly different amounts of computer power to answer.
An AI coding request used to be one chat reply or one line of finished code. In 2026 it can also be a bot that reads your whole project, edits files, runs tests, fails, tries again, and burns compute for a full hour. Those two look identical on your screen. On the company's servers they are nothing alike, and a flat fee cannot cover both.
The whole industry is moving the same way
GitHub put a number on it. Starting June 1, 2026, Copilot stops counting requests and starts counting tokens (a token is roughly three quarters of an English word). Every prompt and answer draws down a monthly allowance, and so does every file the tool reads. Read a big file, pay for a big file.
Github Copilot Moving Usage Based Billing
Cursor felt it first because it rents its models from the big AI makers. In its June 2025 pricing update it said counting requests no longer made sense, because newer models eat far more compute on long jobs.
Anthropic took the slower route. Through 2025 and early 2026 it added weekly caps, then quietly cut limits, then on April 4, 2026 blocked subscription accounts from powering outside coding agents. Users who wanted to keep going had to buy usage bundles or switch to a pay-per-token API key, and some saw costs jump 10 to 50 times.
The Next Web Anthropic Blocks Openclaw Claude Subscriptions Cost
Why the flat plan cannot last
A flat-rate coding plan was a great deal for heavy users and a terrible one for the company selling it. One developer reportedly ran up roughly $15,000 of API usage for about $800 on the top plan. Every time the AI gets cheaper to run, power users just consume more of it, so the gap never closes on its own.
The buildout costs explain the rest. Alphabet reported $35.7 billion in the first quarter of 2026 on data centers and raised its full-year estimate to $180 to $190 billion. Anthropic signed an Amazon deal worth over $100 billion across ten years. Compute is the thing in short supply, and a cheap plan that lets one person run the strongest models for hours fights for the same servers big businesses pay full price to use.
Alphabet Announces First Quarter Results X
What to do now
If a coding tool is part of how you ship work, do not wait for the next plan-change email.
- Set up an API key alongside your subscription. Claude Code supports both at once, so you can run daily work on the subscription and big batch jobs on the key. The official docs walk through it.
- Track your usage so you know whether you are the kind of heavy user these limits target.
- Read the renewal email. If per-token charges get added, you will usually have to opt in, so do not just click through.
Coding helpers are not getting worse. Smaller, cheaper models will handle the boring edits while the powerful ones get saved for planning and hard bugs. AI coding is just turning into a normal metered service, the way you already weigh which cloud server to rent. Reach for the expensive one only when it earns its keep.
