GitHub Copilot Models: Which Ones You Can Pick and Why

GitHub Copilot is not one model. In Copilot Chat, agent mode, Copilot CLI, the GitHub Copilot app and the cloud agent you choose from models made by OpenAI, Anthropic, Google, xAI, Microsoft and Moonshot AI, or pick Auto and let Copilot choose for each request. Which models you see depends on your plan, the tool you are in and, at work, your administrator’s policies. Premium requests are gone: since June 1, 2026 every plan has a monthly allowance of GitHub AI Credits, and what a request costs depends on the model and the tokens it uses. Inline code completions are a separate setting and are not billed in credits on paid plans.

Which models are on the list

GitHub’s supported models page↗ (opens in a new tab) is the list to trust, because it changes often and GitHub says some models “may be replaced or updated over time.” As of September 30, 2026 it lists, by provider:

• OpenAI: the GPT-5 family (GPT-5 mini, GPT-5.3-Codex, GPT-5.4 and its mini and nano versions, GPT-5.5, and GPT-5.6 Luna, Sol and Terra) and the GPT-6 family (GPT-6 Astra, Luna and Sol, and GPT-6.1 Sol).
• Anthropic: Claude Haiku 4.5, Claude Sonnet 4.6, 5 and 5.5, Claude Opus 4.7, 4.8, 5 and 5.5, a fast mode of Opus 4.8 marked as a preview, and Claude Fable 5 and 5.1.
• Google: Gemini 3.5, 3.6, 3.7 and 3.8 Flash.
• xAI: Grok 4.5, 4.6 and 4.7.
• Microsoft: MAI-Code-1.1-Flash. Moonshot AI: Kimi K2.7 Code and Kimi K3.

Three details on that page are easy to miss. GPT-5.3-Codex is the long-term support model and is used if no other model is available. Many models support a 1 million token context window and configurable reasoning levels, and GitHub notes that choosing a larger context window or higher reasoning uses more AI credits. And for Claude Fable 5 and 5.1, Anthropic retains prompts and outputs by default to run safety classifiers; customers can request zero data retention through the end of 2026 under a time-bound exemption. If your company has data-handling rules, read the footnotes before you pick.

Where the model picker is

GitHub’s guide to changing the Chat model↗ (opens in a new tab) puts the picker in roughly the same place everywhere: a dropdown at the bottom of the chat panel.

• GitHub.com: open Copilot from the top right and use the model dropdown under the chat box. You can retry an earlier prompt with a different model.
• VS Code: the dropdown at the bottom of the chat view, which also applies to agent mode. Manage Models adds models from providers such as Anthropic, OpenAI or Gemini with your own key; on Business and Enterprise that needs the “Bring Your Own Language Model Key in Select IDEs” policy.
• Visual Studio, JetBrains IDEs, Eclipse and Xcode: the model dropdown in the chat window. JetBrains adds a Thinking Effort setting for reasoning models.
• Copilot CLI: /model inside a session, or --model when you start one. The choice made with /model is saved to the CLI’s configuration file.
• GitHub Copilot app: a model and a reasoning effort for each session, or Auto. More in the GitHub Copilot app.
• Cloud agent: a picker where you assign an issue to Copilot, mention @copilot in a pull request, or start a task from the agents tab or panel. Where there is no picker, Auto is used.

# pick from the list inside a running session
/model

# or choose when you start
copilot --model <model-name>

Inline completions are a separate setting

The Chat picker does not change the gray suggestions that appear as you type. GitHub’s page on changing the completion model↗ (opens in a new tab) covers VS Code (run “change completions model” from the Command Palette), Visual Studio and JetBrains IDEs, and only when an alternative model is available to you. Code completions and next edit suggestions are not billed in AI credits and are unlimited on paid plans; Copilot Free includes a monthly number of completions.

Auto: letting Copilot choose

GitHub’s page on auto model selection↗ (opens in a new tab) says Auto combines two things: real-time health and availability of each model, and an estimate of how complex your task is. It then routes the request, saving the costlier reasoning models for problems that need them. Auto is available on every plan, in Copilot Chat on GitHub.com, VS Code and JetBrains IDEs, in Copilot CLI, the Copilot app and the cloud agent. In VS Code, the CLI and the app it also has tiers named Efficiency, Balance and Intelligence.

• To see which model answered, hover over the response in Chat and the app; the CLI shows it in the terminal, and the cloud agent at the end of its response.
• Auto never picks a model your plan lacks, one your administrator has turned off, one outside data-residency or FedRAMP restrictions, or an evaluation model you have not enabled.
• On paid plans GitHub discounts model costs when you use Auto in Chat, the CLI, the app or the cloud agent.
• Copilot Free uses Auto only; there is no model choice on that plan.

From premium requests to AI credits

Older guides talk about premium requests and model multipliers, where each chat message to a premium model cost one or more requests. That is the legacy system. GitHub’s billing page for individuals↗ (opens in a new tab) says Chat, the CLI, the cloud agent and third-party coding agents now use AI credits according to the model and the input, output and cached tokens each interaction consumes. A long agent session across a large codebase uses far more than a quick question. The allowance resets at midnight UTC on the first of each month and does not carry over; once it is used, you can set a budget for additional usage, upgrade, or wait. What that looks like when agent mode stops mid-task is in GitHub Copilot agent mode not working.

The best model for GitHub Copilot agent mode

There is no single answer, and GitHub does not give one. Its model comparison page↗ (opens in a new tab) sorts models by the kind of work instead:

• General-purpose coding and writing: GPT-5 mini, GPT-5.3-Codex and GPT-5.6 Terra, for functions, short files, diffs and documentation.
• Fast help with simple or repetitive tasks: GPT-5.6 Luna and Claude Haiku 4.5.
• Deep reasoning and debugging: GPT-5.5, GPT-5.6 Sol, Claude Sonnet 4.6 and Claude Opus 4.7, for bugs across several files, refactoring and architecture.
• Working with visuals such as screenshots and diagrams: GPT-5 mini and Claude Sonnet 4.6.

Note that the comparison page trails the supported models list: it recommends older models than the newest ones you can pick. For agent mode, which reads many files and plans several steps, the deep-reasoning group is the natural starting point. A practical routine is to start on Auto, switch to a reasoning model when a task spans several files or Auto’s result misses, and run the same well-defined task on two models before settling on one for your team. Keep the prompt identical so the comparison means something. How to set the task up well in the first place is in GitHub Copilot agent mode best practices.

What administrators control

• On Copilot Business and Enterprise, the organization has to let members switch models at all.
• Administrators can enable or disable individual models, which also limits what Auto can choose.
• Business and Enterprise have base models, used when no other model is enabled, and long-term support models kept for a year from designation.
• Your own keys in VS Code are governed by a separate policy, and models an organization configures with its own provider keys appear at the bottom of the CLI’s /model list. How Copilot compares with other assistants on admin controls is in GitHub Copilot alternatives for teams.

Recording which model did the work

Once you switch models by task, it is worth knowing afterward which one did what, especially when a change needs revisiting. fenbs, a small task board that Copilot connects to over MCP at https://fenbs.ai/api/mcp, records every change in History under the assistant that made it. Add a line to your Copilot instructions asking the agent to name the model in its comment and to set the task’s test status and test notes to what it actually ran. Each task already has a note for the problem and a plan for the approach, so a model comparison has a fixed brief to work from. fenbs does not track AI credits or model costs, and it has no sprints or due dates; it keeps the list of work and its record.