GitHub Copilot moved all of its plans to usage-based billing on June 1, replacing the old premium-request system with one that meters token consumption, according to GitHub's own announcement. The change ends the flat seat price that made Copilot easy to budget, and a set of smaller coding tools are using the moment to pitch developers who want a bill they can predict.
Under the new model, every plan now ships a monthly allotment of GitHub AI Credits, where one credit equals one cent of model usage. Copilot Pro stays at $10 a month and includes 1,500 credits. Pro+ stays at $39 with 7,000 credits. A new Max plan arrives at $100 with 20,000 credits. Once the allotment runs out, GitHub says each additional credit is charged to the card on file at the same penny-per-credit rate.
For light users, little changes. For anyone running sustained agentic workflows, the math is different, because usage is calculated on input, output and cached tokens at each model's API rate. A developer who leaves an agent running can move from a fixed monthly line item to an open-ended one.
That is the opening the flat-rate tools are aiming at.
"The whole reason people liked a flat fee was that the worst case was knowable," said Santosh Arron, who builds Dropstone, an AI coding CLI that charges one monthly price with no usage overage. "The minute the worst case is whatever your agent did overnight, the product feels different to leave running."
Dropstone charges $15 a month for a plan it says sustains about 450 coding turns a week, with no per-token overage and no five-hour reset window. The company defines a turn as one full exchange with the agent at roughly 15,000 input and 800 output tokens, and says anyone can run the same conversion against a vendor's published quota. It keeps the cost flat, the company says, by running on cache-friendly open-weight models and amortizing the repeated parts of each session, which it measures at above 95% cache-hit rate in sustained use.
The trade-off is the usual one for the cheaper tools. They run on open-weight models rather than a closed lab's flagship. Dropstone says it leads SWE-bench Verified at 91.2%, ahead of Claude Opus 4.7 at 87.6%, but concedes Opus still wins the harder SWE-bench Pro, that GPT-5.5 Pro leads on long-context coding, and that Gemini edges it on multilingual translation.
GitHub's case for the change is that usage-based billing matches what developers actually consume, and that the credit allotments cover typical use. The company also kept Copilot Business at $19 per user and handed existing business and enterprise customers promotional credits for June, July and August to ease the shift.
Whether developers move will depend on how often they hit an overage they did not expect. The Pragmatic Engineer's 2026 survey found that 70% of developers already run two to four AI coding tools at once, so the realistic outcome is not a wholesale switch but a reshuffling of which tool runs the all-day, leave-it-running work. The flat-rate tools are betting that job is now theirs to take.
For developers comparing the options, I wrote a fuller breakdown of how the major plans stack up once you normalize them into the same unit, including where the cheaper ones win and where they still lose: https://blankline.org/research/dropstone-1-5.
Top comments (0)