GitHub Copilot usage limits explained (and how to stop hitting them)

GitHub Copilot doesn't ration everything. Code completions and next-edit suggestions don't count against anything on a paid plan. What's rationed is access to the better chat and agent models. GitHub calls that a premium request, and running out of them is what actually sends people searching for "copilot usage limit" at 11pm mid-refactor.

GitHub has rewritten this system twice in the past year, so what you see in your own account depends on when you signed up and whether you're on a monthly or annual plan. Here's the picture as GitHub's own docs describe it in September 2026, and there are two systems both still live.

The old model: a flat count per month

Until June 1, 2026, Copilot metered premium models by counting requests, with no token math involved.

• Copilot Pro ($10/month): 300 premium requests a month
• Copilot Pro+ ($39/month): 1,500 premium requests a month
• Copilot Free: 50 premium requests and 2,000 code completions a month, tied to a verified GitHub account
• Pro and Pro+ could buy more once they ran out, at $0.04 per extra request
• Running out didn't cut you off. You dropped to one of the included models (GPT-5 mini, GPT-4.1 or GPT-4o) for the rest of the month.

GitHub closed new sign-ups to this system on June 1, 2026. If you're a Pro or Pro+ subscriber who was already on an annual plan before that date, GitHub left you on the request-count system until your plan renews. Everyone else moved to the system below.

The new default: GitHub AI Credits

Since June 1, 2026, Copilot prices a request by which model answered it and how many tokens that took, instead of counting it as one flat unit. One AI Credit is worth $0.01, so a short chat turn on a cheap model costs a fraction of a credit and a long agent session on a reasoning model can cost several.

Code completions and next-edit suggestions still don't touch this budget on any plan. But GitHub's own docs are blunt about what happens once credits run out: additional usage may be capped, and the old habit of quietly dropping to a free model when you hit zero is gone, that cushion belonged to the request-count era. Once you're out, you wait for the monthly reset, upgrade and pay the difference, or buy more usage against a budget you set yourself.

A second ceiling that's easy to miss: session and weekly limits

Separately from credits, GitHub added session limits and weekly limits to stop long, parallel agent runs from spiking the bill. A session limit throttles a single burst of activity. A weekly limit caps total token consumption over a rolling seven days. Hit the weekly limit and Copilot switches you to Auto model selection until the week resets, even if you still have premium requests or credits sitting unused. Pro+ carries a noticeably higher weekly limit than Pro, GitHub puts it at more than five times as much.

What actually burns a request or credit

Copilot CLI can point somewhere else entirely

GitHub built an official way around all of this, but for one surface only. Copilot CLI supports Bring Your Own Key (BYOK): pointing the CLI at any OpenAI Chat Completions-compatible endpoint instead of GitHub's own model providers. That's documented as a Copilot CLI feature specifically. We haven't found the same support documented for the VS Code Copilot Chat extension or the JetBrains plugin, so treat this section as scoped to the CLI.

NeverMetered is one of the endpoints you can point Copilot CLI at. Set three environment variables, COPILOT_PROVIDER_BASE_URL, COPILOT_PROVIDER_API_KEY and COPILOT_MODEL (GitHub fixes those names, not us), and the CLI starts sending its requests to NeverMetered instead of GitHub's model providers. GitHub requires BYOK models to support tool calling and streaming; Qwen3.8-27B supports both. GitHub also recommends at least 128K of context for a BYOK model, and NeverMetered's plans go up to 256K. The full walkthrough is in the Copilot CLI integration guide; this is just the shape of it.

The part that matters here: none of that traffic touches premium requests, AI Credits, or the weekly limit, because it never reaches GitHub's model providers. NeverMetered is priced by parallel streams (how many requests you're running at once), not by token or by request count, so anything routed through Copilot CLI's BYOK setting stops competing with your Copilot budget entirely. Requests past your stream limit wait briefly in a queue instead of failing, and once you pick a plan, its price and stream count hold for 12 months, with 30 days' notice before anything changes. See current plans.