GitHub Copilot imposes multi-day rate limits after fixing token counting bug that undercharged customers

GitHub Copilot customers are reporting rate limits lasting 44 to 181 hours after the company fixed a token counting bug discovered in March 2026 that had been systematically undercounting usage from newer AI models.

The bug affected token counting for models including Anthropic's Claude Opus 4.6 and OpenAI's GPT-5.4, according to Roman Kir, founder of research consultancy StratoAtlas. When GitHub corrected the counting error, configured rate limits snapped back to their intended values, triggering immediate lockouts for customers whose usage patterns had adapted to the artificially low counts.

What GitHub announced

GitHub told customers last week that new limits would be imposed "in the coming weeks" to address "patterns of high concurrency and intense usage" that "place significant strain on our shared infrastructure and operating resources."

The company has taken three immediate actions:

• Imposed new rate limits that can lock users out for multiple days
• Retired Anthropic's Opus 4.6 Fast model for Copilot Pro+ users
• Suspended all GitHub Copilot Pro free trials due to abuse

Customer impact

John Clary, a Copilot Pro Plus hobbyist who reports spending "hundreds of pounds a month on additional credits," hit a 44-hour weekly rate limit. He described being forced to switch to Auto mode, where GitHub selects the model, resulting in "significantly worse performance" as the system appears to favor lower-cost models.

"Auto mode's poor selected model quality frequently taking shortcuts without telling me, which I then have to spend a while getting it to correct," Clary told The Register.

GitHub community discussion threads have received approximately three dozen new complaints in the past two days, with users reporting "obscenely long rate limits" and lockouts lasting several days with "no recovery path" and "no upgrade tier above Pro+."

The economics problem

According to Kir's analysis, the token counting bug masked a fundamental breakdown in GitHub's pricing model. Newer frontier models consume "significantly more infrastructure per request than their predecessors," but GitHub's subscription tiers were priced assuming roughly equivalent costs across premium models.

"The unit of sale – a subscription, a plan tier – had been decoupled from the unit of actual cost," Kir wrote.

GitHub acknowledged in its announcement that the issue stems from "the increased token usage intensity of these newer models."

Industry-wide pattern

GitHub is not alone. Anthropic has implemented similar capacity controls to manage peak demand, and OpenAI Codex users have also complained about rate limits. GitHub suspended free trials citing abuse, suggesting venture capital-subsidized pricing models across AI coding assistants are reaching sustainability limits.

What this means

The GitHub Copilot incident reveals how quickly AI service economics can deteriorate when subscription pricing collides with rapidly escalating inference costs. A token counting bug that likely ran for months allowed usage patterns to form around artificially low costs, then snapped back to economic reality with multi-day lockouts. This suggests other AI subscription services may face similar reckonings as model capabilities and costs continue to scale faster than their fixed-price tier structures can accommodate. The removal of premium models and suspension of trials indicates GitHub is prioritizing cost containment over customer experience.