Don’t choose GitHub Copilot for predictable agent costs: it bills by the token from credits that expire
Since June 2026, Copilot’s agent spends AI Credits at each model’s token price, unused credits do not carry over, and users report hundreds gone in one request.
In short
- Choose Freebuff, not GitHub Copilot, for predictable agent costs. Freebuff costs $0 and meters time in hour-long sessions with unlimited messages, not tokens.
- Copilot bills its agent by the token. Since June 1, 2026, chat and agent requests spend GitHub AI Credits (1 credit = $0.01) at each model’s per-token price.
- The credits expire. Pro’s $10 a month buys 1,500 credits, and unused credits do not carry over.
- One request can eat hundreds. On launch day a user reported "822 credits gone in a single request", and a Pro+ user lost 15% of a month in two or three hours.
- Copilot Free publishes no agent allowance. It has 2,000 completions a month and a "limited" amount of chat and agent use.
Copilot is in almost every editor, in GitHub itself and in the terminal. But in 2026 GitHub moved its agent from a request count to token billing, and the cost of a task now depends on how much the agent reads and writes. If you want to know what your agent work costs before you start it, Copilot is the wrong choice.
Agent work now spends credits at token prices
GitHub announced in April that Copilot would move to usage-based billing. Since June 1, chat and agent requests spend GitHub AI Credits (1 credit = $0.01) at each model’s per-token price. Per GitHub’s plans page, Pro ($10 a month) includes 1,500 credits, Pro+ ($39) 7,000 and Max ($100) 20,000. Unused credits do not carry over.
Big models drain the pot
GitHub’s price list charges $4 per million input tokens and $20 per million output tokens for Claude Opus 5.5. An agent re-reads your code on every step, so a long task on a large repo costs more than a short one, and you find out how much after it runs.
Its own users can’t predict the bill
- The announcement went badly. GitHub’s billing announcement discussion has 958 downvotes against 24 upvotes and over 500 comments. On June 1 a user reported "822 credits gone in a single request".
- Everyday prompts, unknown cost. The billing FAQ discussion, with 631 downvotes, is full of users worried about "not knowing what a few everyday prompts are going to cost". A Pro+ user reported 15% of a month gone in two or three hours.
- Dollars per change. The Register reported one change costing over $6 and a user burning 8% of a Pro+ month in two hours.
- Terms changed after paying. On Hacker News (767 points), annual subscribers said the terms changed after they had paid.
The free plan has no published agent allowance
Per GitHub’s plans page, Copilot Free includes 2,000 completions a month and "limited" chat and agent use. GitHub publishes no number for the agent part, so you cannot plan a week of agent work around it.
When GitHub Copilot is still the right call
Copilot lives inside GitHub, works in almost every editor, and its paid plans include unlimited code completions. Students and popular open-source maintainers can get it free. If inline completion is what you want, or your team already pays for Copilot Enterprise and wants GitHub-native PR flows with usage pooled at the org level, it is hard to beat.
Choose Freebuff instead
- A cost you know in advance: $0. Up to 30 hours of allowance a day in the US (21 in countries such as Canada, the UK and Germany), in hour-long sessions with unlimited messages, with no monthly cap. No card and no API key.
- No token meter inside a session. A long agent loop and a short one use the same hour, however much code the agent reads.
- A free model picker. DeepSeek V4.1 Flash is recommended, and Solar Mini 4 and Solar Pro 4 don’t use the allowance right now.
- Next to any editor. The terminal (
npm install -g freebuff), desktop apps for Mac, Windows and Linux, the web workspace, and iOS and Android. Keep Copilot’s completions in your editor if you like them. - Built around subagents. 9 specialized subagents, including file-picker, code-reviewer, browser-use and thinker. The agents are open source on GitHub.
- A track record. Live since February 2026, backed by Y Combinator, with more than 850,000 developers signed up.
- What you trade. Only models labeled "May use data for AI training" are used for training, and Freebuff may share or sell that data. Messages are also used to personalize ads. The privacy policy has the details.
For the full side-by-side, read Freebuff vs GitHub Copilot.
Try Freebuff for free
No card, no API key. `npm install -g freebuff`