Microsoft's GitHub last week told Copilot customers that they'd have to reduce their use of the AI service to ease the strain on company servers. This follows the company's discovery last month of a token counting bug that appears to have broken the company's pricing model.
"As GitHub Copilot continues to rapidly grow, we continue to observe an increase in patterns of high concurrency and intense usage," GitHub said. "While we understand this can be driven by legitimate workflows, this type of usage places significant strain on our shared infrastructure and operating resources."
To improve the Copilot experience, the company said, limits would be imposed in the coming weeks "to better balance capacity" and to improve "overall service reliability."
GitHub also said that it was retiring Anthropic's Opus 4.6 Fast for Copilot Pro+ users.
Anthropic has run into similar capacity problems and has also taken steps to steer consumption patterns by discouraging usage during peak demand amid growing complaints from developers. Developers using OpenAI's Codex have also been objecting to rate limits. The bill for the venture capital-fueled all-you-can-eat token buffet has come due.
In a further cost-cutting move, GitHub has suspended all GitHub Copilot Pro free trials due to abuse.
GitHub Copilot customers who have encountered unexpected rate limits are not pleased.
"I use Github Copilot Pro Plus as a hobbyist and frequently spend hundreds of pounds a month on additional credits," John Clary told The Register in an email. "I've been frequently receiving short rate limits interrupting the request with no listed countdown, only to try again in X seconds, with X being under 30 seconds or so. However, I usually wait roughly 15 minutes to resume as, if you don't, you'll receive another right when the AI is working on something important."
Clary said that he was able to manage that, but Monday evening he hit a 44-hour "weekly rate limit," which he believes is related to GitHub's announcement last week.
Initially, he assumed that it was a bug, and waited a while, but realized after an hour that the rate limit barrier would not be dropped. So he opened a support ticket and switched reluctantly to Auto mode, where Copilot selects the model on the user's behalf.
Presumably Auto mode favors models with lower inference costs because Clary said it offered significantly worse performance.
"Since then, little progress on my projects has been made, with 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," he said.
Similar sentiment has surfaced in GitHub Copilot community discussions. Long-running threads about rate limits have seen a surge in complaints – about three dozen in the past two days for this one discussion.
Customers have raised the alarm about what they characterize as "obscenely long rate limits," with people reporting that they've been told to wait several days before they can use the service again.
GitHub has framed the issue as a matter of server capacity, but it also looks like an attempt to throttle unanticipated costs.
Roman Kir, founder of research consultancy StratoAtlas, said in an online post that GitHub has had to impose limits because its pricing model broke.
"In March 2026, GitHub discovered a bug: its rate-limiting system had been undercounting tokens from newer models like Claude Opus 4.6 and GPT-5.4," he wrote. "These models consumed significantly more infrastructure per request than their predecessors — but the bug made them look cheap. When GitHub fixed the bug, limits snapped back to their configured values. For users with normal workflows, those values were suddenly too tight. 181-hour lockouts. Coding agents killed mid-task. No recovery path. No upgrade tier above Pro+."
According to Kir, GitHub's capacity issues follow from the collapse of its pricing structure. It used to be that prompts to premium models incurred roughly the same infrastructure cost. With the latest frontier models, he said, that's no longer the case.
"The unit of sale – a subscription, a plan tier – had been decoupled from the unit of actual cost," he wrote.
This is essentially what GitHub said when it acknowledged that its bug fix broke expected usage patterns as a result of "the increased token usage intensity of these newer models."
An individual posting under the name Delsin-Yu to one of the rate limiting discussions speculated as much.
"GitHub's level of subsidization appears to have reached a point that is no longer sustainable," said Delsin-Yu. "As a result, this feels less like a bug and more like a business decision. …"
GitHub did not respond to a request for comment. ®
Source: The register