You can now choose how much reasoning effort a task spends, alongside the coding agent and the model it runs on.
Per agent, per run. An agent carries an effort the way it carries a model, and every task it starts runs at that effort. Create a task over the API or the CLI and you can pass an effort for that run alone, so a job that needs more thinking gets it without changing what the agent does tomorrow.
The effort on the request wins. Below that, the agent's own setting carries over only when the run uses the coding agent that setting belongs to, since the levels are each tool's own words. Switch coding agents for a run and it starts from that tool's default.
Per-tool levels. Claude Code runs low, medium or high. Codex adds minimal at the bottom and extra high at the top. The app reads the levels from the server, so you are offered exactly what the tool you picked accepts, and an effort a tool does not recognize falls back to its default rather than failing the task.
Per-turn details. Every finished agent response carries a details card naming the coding agent, the model and the effort it ran at, plus the subscription seat it was billed to when it ran on one. A turn with no pinned effort still ran at one, so the card shows the coding agent's default and marks it as the default. The composer names the computer and the model before you send, and the effort as well when the agent pins one.
New models, same day. The model picker reads its list from the server, so a model added there shows up the same day, with no app update. Opus 5, GPT-5.6 Sol and GPT-6 Astra all reached the picker that way. Each new model is added below the current default, so agents keep the model they already run until you choose otherwise.
Set an agent's effort
Open the agent, go to Configure, expand Advanced settings, and set Effort, just below Model. Start at the default and move a level when the work asks for it: higher effort on a design or a gnarly bug, lower on a rename that just needs doing.
Read the models docs for each coding agent's levels, the order of precedence, and the effort field on task create.