GitHub repriced Copilot per token after agent usage doubled its inference bill

In April 2026, GitHub announced that from June 1 it would shift Copilot from monthly allocations of 'requests' and 'premium requests' to usage-based billing, pitched as a way to 'better align pricing with actual usage'. The company's framing: 'Today, a quick chat question and a multi-hour autonomous coding session can cost the user the same amount', and having 'absorbed much of the escalating inference cost behind that usage' was 'no longer sustainable'.

Under the new system, subscribers receive a monthly allotment of 'AI Credits' matching their subscription payment; beyond that, pricing is calculated on token consumption — input, output and cached tokens — at each model's listed API rates, which vary widely: OpenAI's GPT-5.4 Mini at $4.50 per million output tokens versus $30 for GPT-5.5. Simple code suggestions and Next Edit stay credit-free, while Copilot code reviews will cost GitHub Actions minutes. A 'preview bill' tool let users forecast their charges before the change took effect.

The repricing capped a rough stretch: days earlier GitHub paused new signups, tightened usage limits and removed Anthropic's Claude Opus models from lower-tier Pro plans. AI critic Ed Zitron cited leaked internal documents showing week-over-week Copilot costs had nearly doubled since January, timed to the rise of nearly always-on multi-agent assistants like Openclaw. Anthropic made parallel moves — charging large Claude Enterprise subscribers full compute cost, briefly testing Claude Code's removal from the $20 Pro plan, and adjusting limits during 5-11am Pacific peak hours.

AI Credits matched to the subscription cover the baseline; overage is billed per input, output and cached token at each model's listed API rate.

Flat-rate premium requests made a one-line chat and a multi-hour autonomous agent session cost the same, so the heaviest users could burn unbounded compute.

Agentic assistants ran near-continuously, nearly doubling weekly inference costs and making the subsidy unsustainable exactly for the most engaged customers.

Per-token API rates already existed, so realignment meant mapping Copilot usage onto prices the market already pays instead of inventing a new currency.

All-you-can-eat pricing hides a marginal cost somebody pays. When one user class discovers the buffet, meter the expensive tier at cost and leave the boring 90% of requests free.

GitHub positioned the change as 'reducing the need to gate heavy users' rather than throttling them, and gave subscribers a preview-bill tool ahead of the June 1, 2026 switch. The report expects such moves to spread as AI companies chase profits that 'have so far been illusory' amid a compute shortage.