GitHub Adds Reasoning-Level Controls to Copilot Cloud Agent
Paid-plan users can choose how much supported models reason for each delegated task, trading potential quality gains against token and credit use.
GitHub announced on August 3 that Copilot cloud agent users can now choose a reasoning level when delegating a task to a model that supports the control. The setting is available across paid Copilot plans that include the cloud agent.
What changed
The new selector appears alongside the model choice when a user starts a cloud-agent task. GitHub says the selected level governs how much the model reasons before responding, and the cloud agent applies that setting for the run.
This makes reasoning effort an explicit task-level input instead of leaving the entire choice inside the provider's model configuration. A developer can choose a lower level for routine work or assign more reasoning to a complex task, provided the selected model supports the option.
The cost trade-off
GitHub says a higher reasoning level can improve answers to complex problems, but it also consumes more tokens and therefore more Copilot credits. The control gives users a visible way to make that trade-off before a delegated run begins.
For individual developers, the practical question is whether extra reasoning materially improves a task enough to justify the additional credits. For teams, the selector also creates a governance choice: administrators and engineering leads may need guidance on when higher settings are appropriate so that difficult jobs receive enough model effort without turning every routine task into a higher-cost run.
Availability and limits
GitHub lists the control for paid Copilot plans that include cloud agent, covering Pro, Pro+, Business, Enterprise and Max. Support still depends on the model, so the selector does not imply that every model exposes multiple reasoning levels.
The one-minute changelog does not quantify quality gains, token multipliers or credit costs for the available levels. It also does not provide task-level benchmarks comparing the settings. Users therefore have the control before they have public evidence describing how much it changes outcomes or spend across different kinds of coding work.
Status
Confirmed. Internal confidence is medium because availability and behaviour are documented in GitHub's official changelog, while the performance and cost trade-offs remain qualitative vendor guidance rather than independently measured results.
Sources
Update note: Last reviewed 2026-08-05. We will revise this post if GitHub publishes measured quality, token or credit differences between reasoning levels.
Sources
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.