Agentic coding workflows are now routinely generating costs that exceed what users pay per month. GitHub’s response, pausing new sign-ups for Pro, Pro+, and Student plans and tightening usage caps, signals that the era of unlimited AI assistance at fixed prices is ending.

GitHub has paused new sign-ups for its Copilot Pro, Pro+, and Student plans and tightened usage limits across all individual tiers, citing a fundamental mismatch between how developers now use the product and the infrastructure it was built to support.

The company’s VP of product, Joe Binder, said in a blog post that agentic coding workflows, long-running, parallelised sessions in which AI agents and subagents tackle complex problems autonomously over extended periods, are now routinely consuming more compute than users pay for in a month.

“It’s now common for a handful of requests to incur costs that exceed the plan price,” Binder wrote.

The change, effective 20 April, leaves Copilot Free as the only plan still accepting new individual sign-ups. Existing users retain access to their current plans and can upgrade between tiers, but GitHub has given no timeline for resuming new subscriptions.

Pro and Pro+ subscribers who contact GitHub support between 20 April and 20 May can cancel and receive a refund, with no charge for April.

The usage changes that accompany the pause are structured to push heavier users towards the pricier Pro+ tier. GitHub is tightening both session and weekly token limits on individual plans, caps that govern how many tokens a user can consume in a given time window, separate from the premium request entitlements that determine model access.

A user can have premium requests remaining and still hit a usage limit because the two systems operate independently. Pro+, at $39 per month, now offers more than five times the limits of the $10-per-month Pro plan.

Usage warnings are being added to VS Code and the Copilot CLI so developers can see approaching limits before hitting them mid-workflow.

Model access is also being restructured. Opus models, Anthropic’s heaviest and most capable models, are being removed from the Pro plan entirely.

Opus 4.7 remains available on Pro+. Opus 4.5 and 4.6, previously announced for removal from Pro+, are being removed from that tier as well. The pattern is straightforward: the most compute-intensive models are migrating exclusively to the most expensive individual tier.

The economics behind the move are unusually candid for a Microsoft product announcement. Copilot was originally designed for code completion, short, stateless suggestions that consume modest compute per interaction.

Agentic coding, by contrast, involves sessions that can run for hours, spawn multiple parallel threads, and generate token volumes that bear no resemblance to the autocomplete interactions that shaped the original pricing structure.

GitHub’s own Copilot features, including the /fleet command for parallel workflows, are now listed among the behaviours GitHub is asking its own users to limit.

This is not the first sign of strain. The week before the sign-up pause, GitHub had already suspended Copilot Pro free trials due to abuse, a narrower measure that hinted at the broader capacity pressure to come.

And the sign-up pause itself arrives at a politically awkward moment for GitHub with its developer user base.

In late March, the platform came under significant backlash after developers discovered that Copilot had been inserting promotional “tips”, including an advertisement for productivity app Raycast, into pull requests, in some cases appearing as if written by the developer rather than the AI.

The feature was disabled the same day, with GitHub’s VP of developer relations, Martin Woodward, saying the behaviour had become “icky” after Copilot’s reach was extended to pull requests it hadn’t created.

GitHub described it as a programming logic issue, not an advertising strategy. More than 11,000 pull requests were affected before the rollback.

The broader pattern, analysts say, is structural. Charlie Dai, vice president and principal analyst at Forrester, said the move shows how agent-driven coding is shifting workloads towards longer-running and parallel sessions that create higher and less predictable compute demand.

“Cost structures built for lightweight assistance no longer hold,” Dai said, “and this puts pressure on GPU capacity, reliability, and unit economics.” He added that similar usage restrictions from major model providers suggest capacity rationing is likely to become a feature of the industry as agentic development becomes routine.

For enterprise engineering leaders, Dai said the episode is a reminder to evaluate AI coding tools as metered infrastructure rather than unlimited productivity layers.

Faisal Kawoosa, founder and chief analyst at Techarc, said the dynamic is a familiar one. “First you give users access to a tool with relatively open usage, and then gradually start defining limits as adoption grows,” he said.

Kawoosa added that the next step is likely to be more differentiated plans that create clearer monetisation opportunities, noting that GitHub’s depth of integration into developer workflows gives it unusual leverage: “a developer can live without an email ID, but not a GitHub account.”

Whether competitors, including Claude Code, Cursor, and Codeium, can move quickly enough to absorb frustrated Copilot users before GitHub recalibrates its pricing structure is the open question the market is now watching.