GitHub will start charging Copilot users by the token on June 1, ending the flat-rate model that has subsidized the service's heaviest agentic users. The Microsoft-owned company said in its announcement that week-over-week inference costs have outpaced what subscription revenue can cover, and that a quick chat question and a multi-hour autonomous coding session currently cost the user the same amount. Subscribers will still get a monthly allotment of AI Credits matching their subscription fee, with anything beyond that billed at API rates.
Those rates vary sharply by model. OpenAI's GPT-5.4 Mini runs $4.50 per million output tokens, while GPT-5.5 runs $30 per million — a nearly 7x spread that will reward users who route routine work to cheaper models. Token consumption per prompt also varies with how much reasoning a model performs, meaning a single request to a high-end model can quietly burn through a meaningful slice of a monthly credit balance.
AI critic Ed Zitron reported last week, citing leaked internal documents, that week-over-week Copilot costs have nearly doubled since January 2026. That trajectory tracks the rise of agentic coding assistants like Openclaw, which run nearly always-on multi-agent workflows and consume tokens at a rate flat-fee subscriptions were never priced for.
Key facts
- 01GitHub Copilot moves to usage-based billing on June 1, 2026, with monthly AI Credits sized to match each user's subscription fee.
- 02Overage pricing follows API rates: $4.50 per million output tokens for GPT-5.4 Mini, up to $30 per million for GPT-5.5.
- 03Leaked documents cited by Ed Zitron show week-over-week Copilot costs have nearly doubled since January 2026.
- 04GitHub last week paused new signups, tightened limits, and pulled Claude Opus from lower-tier Pro plans.
- 05Anthropic is making parallel moves, briefly testing removal of Claude Code from its $20-per-month Pro plan.
GitHub framed the move as a sustainability measure rather than a price hike. The company said it has "absorbed much of the escalating inference cost behind that usage," but that lumping all premium requests into a single bucket "is no longer sustainable." The new structure, GitHub said, "reduces the need to gate heavy users" who currently exploit the flat-rate plan.
“GitHub says week-over-week Copilot costs have nearly doubled since January, and that lumping a quick chat and a multi-hour agent run into the same 'premium request' bucket is no longer sustainable.”— Jaeden Schafer
Simple completions and Next Edit suggestions will remain free of credit consumption. Copilot code reviews, however, will now draw down GitHub Actions minutes, adding a second metered dimension to the bill. A preview bill tool will let subscribers forecast what their current behavior would cost under the new model before June 1.
The pricing reset comes after a turbulent week at GitHub. The company paused new Copilot signups, tightened usage limits, and pulled Anthropic's Claude Opus models from lower-tier Pro plans, saying the moves were "necessary to ensure we can serve existing customers with a predictable experience." Read together with the June 1 changes, the picture is of a service running hot against its compute budget.
Anthropic is making the same calculation. The Information reported the company has begun charging large Claude Enterprise customers for the full cost of compute rather than discounting tokens through subscriptions. Anthropic also briefly tested removing the compute-heavy Claude Code from its $20-per-month Claude Pro plan, and has been throttling usage during the 5 am to 11 am Pacific peak window to protect reliability.
The pattern lines up with prior AI Chat Daily coverage of Red Hat's Tank OS work to constrain Openclaw deployments — agentic tooling is no longer a research curiosity but a meaningful cost line. Always-on agents that loop through code, tests, and edits behave nothing like the chat-style usage these subscriptions were originally priced against.
There is a real risk in GitHub's approach. Token-metered billing makes costs unpredictable for the individual developer, and that unpredictability tends to chill experimentation — exactly the behavior Copilot was supposed to encourage. Teams that have built workflows around "unlimited" agent runs will need to add cost monitoring, and finance leaders will need to model usage variance the way they model cloud spend. Some heavy users may simply route to a competitor that still subsidizes the workload, at least until that competitor blinks too.
GitHub said the change is "designed to deliver a more sustainable and reliable product experience by aligning pricing to actual usage and costs." That phrasing concedes the point: subscription-subsidized AI was a customer-acquisition strategy, not a business model.
The broader read is that the era of flat-rate AI coding tools is closing. With OpenAI, Anthropic, and now GitHub all repricing toward consumption, the developer-tools market is converging on the same metered model that defines public cloud — and the economics of agentic AI will increasingly look less like SaaS and more like AWS. For Microsoft, the question is whether Copilot's brand strength can hold paying users through a transition that, for the loudest power users, will feel like a tax.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
