GitHub Copilot Now Lets You Pick Claude or GPT-5 for Agent Tasks

On April 14, 2026, GitHub added model selection for Claude and Codex agents on github.com. Until now, when you kicked off an agent task in GitHub Copilot — like having the agent write code for an issue — GitHub picked the model for you. Now you choose. The options are Claude Sonnet 4.6, Claude Opus 4.6, GPT-5.2-Codex, GPT-5.3-Codex, and GPT-5.4.

This is a meaningful shift. It makes GitHub Copilot a model-agnostic platform rather than just a gateway to OpenAI's models. For teams with preferences about which AI handles their code, this matters.

How to Access Model Selection

Selecting a model for agent tasks on github.com:
1. Open any GitHub issue or pull request
2. Look for the Copilot icon in the sidebar (or the 'Ask Copilot' button)
3. Click 'Start agent task' or equivalent
4. Before the agent begins, you'll see a model picker dropdown
5. Available options:
   - Claude Sonnet 4.6 (fast, strong on complex reasoning)
   - Claude Opus 4.6 (most capable Claude, slower)
   - GPT-5.2-Codex (OpenAI's specialized coding model, fast)
   - GPT-5.3-Codex (improved version of 5.2, strong on benchmarks)
   - GPT-5.4 (latest GPT-5 variant, broad capability)
6. Select your model, then describe the task in the prompt field
7. The agent runs asynchronously — you get a notification when done

Which Model to Pick for What

The honest answer is: try a few and see which feels right for your codebase. But based on what each model is optimized for, here are practical starting points.

Model selection guide for Copilot agent tasks:

Claude Sonnet 4.6 — Best default for most tasks
- Multi-file refactors that need reasoning across the whole codebase
- Writing tests that handle edge cases correctly
- Bug fixes where understanding context matters more than raw speed
- Any task where you want the agent to explain its decisions

Claude Opus 4.6 — Use when quality matters most
- Architecture-level changes that will be hard to reverse
- Security-sensitive code paths
- Code review of a large PR before it merges to main
- Tasks where a wrong decision costs significant cleanup time

GPT-5.3-Codex / GPT-5.4 — Fast, high-quality output
- Generating boilerplate that follows existing patterns
- Well-scoped feature additions with clear specs
- Documentation generation
- Tasks where you want strong performance on SWE-Bench style problems

Asynchronous Agent Tasks on github.com

This feature is specifically for agent tasks on github.com — not VS Code or the CLI. When you start an agent task from a GitHub issue, the agent spins up, works in a sandboxed environment, and creates a draft PR with its changes. You review the PR like any other pull request. The agent doesn't have access to your local machine — it works from the code in the repository.

This is actually a meaningful workflow change for teams: you can assign issues to Copilot directly from the GitHub interface, with no developer touching the task until they review the resulting PR. Model selection makes that workflow more flexible — you can send complex architecture issues to Opus and fast boilerplate tasks to Codex.

What This Means for Teams

GitHub is positioning Copilot as a model-agnostic development platform. You're not locked into OpenAI's models anymore. For teams that have found Claude performs better on certain types of tasks — and many have — this is the option they've been asking for. It also means your Copilot Enterprise subscription now gives you access to Anthropic's frontier models without a separate API contract.