GitHub lets Claude and Codex agents pick newer models per task

GitHub is quietly turning third-party coding agents from one-model products into configurable workflows. In its April 14 changelog, the company says users of Claude and Codex coding agents on github.com can now choose the model they want when starting a task. That sounds minor until you remember how much coding-agent behavior shifts with model choice: latency, reasoning depth, cost, and patch quality can all swing depending on whether a task needs speed, broader context handling, or more aggressive planning.

The model menu is not symbolic. GitHub says Claude coding agent users can currently pick among Claude Sonnet 4.6, Claude Opus 4.6, Claude Sonnet 4.5, and Claude Opus 4.5. Codex users can choose GPT-5.2-Codex, GPT-5.3-Codex, or GPT-5.4. In other words, GitHub is exposing a seven-model surface across Anthropic and OpenAI rather than forcing teams into a single default. That makes the agent tab look a little more like an orchestration layer and a little less like a branded wrapper around one backend.

This matters because coding agents are increasingly judged by operational fit, not only raw benchmark reputation. A team debugging a stubborn infrastructure issue may want a slower, stronger model. A team burning through repetitive pull-request chores may prefer a faster, cheaper option. By moving model choice into task kickoff, GitHub is acknowledging that developers do not experience the agent as a monolith. They experience a stack of trade-offs. The UI now lets them express that trade-off directly instead of routing every task through the same model profile.

There are still limits. Access is bundled with an existing Copilot subscription rather than being a universal default, and GitHub says Copilot Business and Enterprise customers need the relevant Anthropic Claude or OpenAI Codex policy enabled by an administrator. The repository owner or organization also has to enable the agent from Settings > Copilot > Cloud agent. So the feature opens flexibility, but inside managed guardrails rather than free-form experimentation.

The broader signal is that github.com is becoming a model marketplace interface as much as a coding surface. Once users can choose among multiple Anthropic and OpenAI models at task start, expectations change fast: teams begin to compare not just agents but routing logic, pricing, governance, and which new models appear first. GitHub is effectively betting that coding-agent competition will be won by giving developers sharper control over when to trade speed for depth, not by pretending one model is always right for every repository.