How much does AI cost? GitHub Copilot users adopt a new pricing model.

“Even though I was very careful the first day, testing and using a little bit, it still ate up 840 records,” one user said. he wrote testing Claude Sonnet 4.6 via Copilot today. “I haven’t even done the hardest work,” one user later complained report use represents 21 percent of their monthly Pro Copilot credit in one day. “I feel like I’m going somewhere soon.”

Spending all of your organization’s monthly AI budget of 8,000 in one day… probably isn’t sustainable.

Amidst the price changes, many GitHub Copilot users are predicting and publicly threatening to cancel their subscriptions or seek alternative AI coding solutions. But others say they can get used to the new world of trees. By Henri Kinnunen they write that he only burned 161 records in the “production day” using Claude 5.3-Codex through Copilot, due to the restraint of “deliberate modification by AI.” In Bluesky, by Neil Hewitt wisely realized that continuing to chat for three days on Copilot probably wasn’t smart now, because it meant sending “the entire chat history as news all the time… Hey, tokens use history… it’s not rocket science.”

While some Copilot users are jumping ship for other services with more generous usage limits, this type of customer support approach could soon drive Copilot’s pricing across the industry. When this happens, LLMs that work well with their marks can win the financial battle; on Reddit, one user is previous discussion how they integrated Deepseek into their GitHub VSCode repository at a cost of “about 7 cents for 15 million tokens.” While you may say “you get what you pay for,” some AI users are now thinking about a world where they also have to pay for what they get.