Image: GitHub Copilot's Metered Billing Went Live Yesterday. The Numbers Are Coming In.
GitHub Copilot's switch to usage-based billing hit on June 1. By June 2, developers were sharing real-world credit burn rates — and some are already planning exits.
GitHub Copilot’s new usage-based billing went live June 1. Within 24 hours, developers started posting actual credit burn numbers. The picture isn’t good for heavy users.
The Register reported several specific cases on June 2:
- A Pro+ subscriber ($39/month, 3,900 credits included) burned through 8% of their monthly allocation in two hours.
- One developer spent over $6 on a single feature request.
- Another consumed 1,180 credits — roughly 16% of the Pro+ monthly pool — on what they described as “mediocre suggestions that didn’t solve their problem.”
- A fourth developer calculated that their 7,000-unit quota would run out in less than two days at their normal usage rate.
These are individual cases, not averages. But they’re the kind of numbers that shape perception.
What’s happening mechanically
Code completions and Next Edit suggestions remain unlimited — they don’t touch your credit balance. Everything else does: chat requests, agentic sessions, multi-step Workspace tasks. The longer and more complex the session, the more tokens, the faster the credits go.
The billing is token-based at each model’s API rate. GPT-4.1 costs more per session than a smaller model. An agentic session that opens files, runs tests, and iterates over several steps can accumulate token costs that would have been invisible under the old flat-rate model.
What developers are doing about it
The migration alternatives being discussed on Reddit and in GitHub’s own community forums: direct API access through Anthropic or OpenAI, RooCode (open-source VS Code extension with BYOK), LM Studio for local models, and OpenRouter for model breadth and the fact that unused API credits don’t expire.
Microsoft’s response is to point developers toward spending controls. Business and Enterprise accounts can set credit limits at the organization, team, or individual level. There’s also a model selection tool that lets you route requests to cheaper models to conserve credits.
The company has also positioned “Copilot Max” as a higher-capacity tier for power users — essentially an escape valve for developers whose workloads genuinely require sustained agentic sessions.
The honest framing
The backlash has been loudest from developers who used Copilot heavily for long chat sessions and agentic workflows. Developers who primarily use code completions are largely unaffected — that part stayed unlimited.
The problem is that GitHub spent the last 18 months positioning Copilot Workspace and agentic features as the future of the product. It got developers building workflows around multi-step AI sessions. Now those workflows are the expensive ones.
The billing model change isn’t irrational — running long agentic sessions on frontier models is genuinely more expensive than serving autocomplete tokens. But the timing, arriving the same week as the Project Polaris announcement at Build 2026 and Copilot Workspace going GA, means developers are being asked to pay more for features they’re simultaneously being sold on using more heavily.
Sources: The Register (O’Ryan Johnson, June 2, 2026)