GitHub halts new sign-ups for Copilot as autonomous AI disrupts the economics.

      Agentic coding workflows are increasingly resulting in costs that surpass the monthly fees users pay. GitHub's recent decision to halt new sign-ups for its Pro, Pro+, and Student plans, along with imposing stricter usage limits, indicates that the age of unlimited AI support at fixed rates is coming to an end.

      The company has announced a suspension of new sign-ups for its Copilot Pro, Pro+, and Student plans and is tightening usage restrictions across all individual tiers. This change is due to a core discrepancy between developers' current use of the product and the infrastructure it was originally designed to accommodate.

      In a blog post, Joe Binder, GitHub’s VP of product, mentioned that agentic coding workflows—extended, parallel sessions where AI agents autonomously address complex issues—are now routinely utilizing more compute resources than what users pay for monthly. "It's common for just a few requests to generate costs that exceed the plan's price,” Binder noted.

      Effective April 20, Copilot Free remains the sole plan allowing new individual sign-ups. Current users will keep access to their existing plans and have the option to upgrade, but GitHub has not provided a timeline for the resumption of new subscriptions.

      Pro and Pro+ users who reach out to GitHub support between April 20 and May 20 may cancel their subscriptions and obtain a refund, with no charge for the month of April.

      The adjustments in usage are designed to encourage heavier users to migrate to the more expensive Pro+ tier. GitHub is tightening both session and weekly token limits on individual plans, which dictate the number of tokens a user can use within a specified timeframe, and these limitations are separate from the premium request entitlements that determine model access.

      A user can still hit a usage cap even if they have remaining premium requests, as the two systems function independently. Pro+, now priced at $39 per month, provides over five times the limits of the $10-per-month Pro plan.

      Usage notifications will be added to VS Code and the Copilot CLI, allowing developers to be alerted about approaching limits before they exceed them during their workflow.

      Access to models is also being restructured, with Opus models, Anthropic’s most powerful models, being entirely removed from the Pro plan. Opus 4.7 will remain available on Pro+, while Opus 4.5 and 4.6—previously slated for removal from Pro+—will also be taken out from that tier. The trend is clear: the highest compute-demand models are being shifted exclusively to the most expensive individual tier.

      The reasoning behind this shift is notably transparent for a Microsoft product announcement. Copilot was initially intended for code completion, providing brief, state-independent suggestions that require less compute power per interaction. In contrast, agentic coding involves sessions that can persist for hours, generate multiple parallel threads, and produce token volumes vastly different from the autocomplete interactions that influenced the original pricing model.

      GitHub's own Copilot features, such as the /fleet command for parallel workflows, are now listed among the activities GitHub urges its users to limit.

      This situation is not the first indication of strain. Just a week prior to the pause in new sign-ups, GitHub had already suspended Copilot Pro free trials due to misuse, a smaller measure reflecting the larger capacity pressures at play.

      Additionally, the timing of the sign-up pause is politically sensitive for GitHub among its developer community. In late March, the platform faced backlash after developers found that Copilot was injecting promotional “tips,” including ads for productivity app Raycast, into pull requests, sometimes appearing as if they had been authored by the developer instead of the AI.

      This feature was disabled that same day, with GitHub’s VP of developer relations, Martin Woodward, stating that the behavior became “icky” after Copilot’s capability extended to pull requests it did not originate. GitHub characterized it as a programming logic issue rather than a marketing tactic, affecting over 11,000 pull requests before the rollback.

      Analysts view this as indicative of a broader trend. Charlie Dai, vice president and principal analyst at Forrester, commented that this change reflects how agent-driven coding is shifting workloads toward longer, parallel sessions that generate higher and more unpredictable compute demands.

      “Cost models designed for lightweight assistance are no longer sustainable,” Dai noted, “and this places pressure on GPU capacity, reliability, and unit economics.” He further mentioned that similar restrictions from major model providers suggest that capacity rationing is likely to become commonplace in the industry as agentic development becomes standard.

      For enterprise engineering leaders, Dai stated this incident serves as a reminder to assess AI coding tools as metered infrastructure rather than limitless productivity resources.

      Faisal Kawoosa, founder and chief analyst at Techarc, observed that this dynamic is familiar. “First, you empower users with a tool that has relatively open usage, and then gradually start establishing limits as adoption escalates,”