The enormous amount of money shuffling through AI's circular economy has created a money pit, and the beast therein demands sacrifices. Not too long after Anthropic shocked most of the developer world by bumping up its prices and moving Claude Code to the Max plan, GitHub Copilot has followed suit. The move from subscription to usage-based billing has left many users looking down the barrel of massive bills, according to an Ars Technica report.
An ongoing discussion on the GitHub community forums includes plenty of customer testimonials, as do X posts from many different users. The overall gist is simple: many users are reporting that their bills would increase by several orders of magnitude, or that the limit is so low that a subscription plan is now either extremely limited or useless. There's even a community cost estimator that popped up a while back when the news first came to light.
To wit, the AI allowance in each of GitHub's subscription plans has three tiers: the $10 Pro plan gets you 1,500 credits; the $39 Pro+ plan contains 7,000; and the $100 Max subscription nets you 20,000 credits. While it's good for Microsoft to specify precisely how many tokens each plan includes, it's worth noting that in obvious cases like long-running conversations or queries on large projects, it's exceedingly hard to estimate how many tokens any given query will use, as shown by data from a 2025 paper.
Unsurprisingly, some users are reporting that even light usage, or being "super cautious", they went through significant chunks of their monthly allotment in the blink of an eye. Switching the underlying model for a query can drastically change the calculations; for example, using Claude Opus or GPT will be far pricier than using Gemini Flash. Some subscribers also warn fellow vibe-coders about resurfacing long-running conversations, as the nature of an AI bot requires the entire conversation to be re-sent again and again, quickly chewing through usage limits.
Besides being more judicious about selecting the right model for a given query, some folks are experimenting with harnesses to use token-efficient models like DeepSeek to keep costs down. Those using an AI bot as an agent also have to exercise extreme caution, as there's no shortage of reports and horror stories about clankers left unchecked, racking up gigantic bills.
Although there's no shortage of vitriol directed at Microsoft, as there was when Anthropic bumped up Claude's pricing, the harsh reality is that investors are growing tired of dumping money into AI companies and want to see cash inflows that are at least in the same galaxy as their expenditure.
Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.
