Deep Reasoning

Deep Reasoning is a multi-step model-calling pattern that breaks complex tasks into explicit phases: planning, drafting, critiquing, revising, and finalizing. Instead of relying on a single model completion to produce a good answer, the platform orchestrates multiple focused calls and lets the model review and improve its own work.

A long system prompt asks the model to plan, write, and review all in one completion. Deep Reasoning runs each phase as a separate model call with dedicated instructions. This produces better results because each call has a focused objective, and intermediate outputs are inspectable and auditable.

Yes. Every step produces a trace artifact — a persisted, inspectable document. You can view the plan, read the critique, compare the draft to the revision, and audit the entire reasoning chain. These are real documents, not ephemeral logs.

The job stops at the current step boundary and records a budget-exceeded error. All completed steps and their trace artifacts are preserved. You can inspect the partial results and optionally re-submit with a larger budget.

Yes. Request cancellation via the API or dashboard. The executor checks for cancellation between steps and stops cooperatively. All completed work is preserved and the invocation receipt shows which steps finished.

Each model call within the reasoning pipeline is metered individually — LLM tokens plus the standard platform premium. The total cost of a reasoning job is the sum of its step costs. Budget controls ensure you never exceed your configured ceiling.