Reasoning & hosted tools

In Simple mode, the composer shows reasoning and hosted-tool controls only when the selected route supports them. These settings apply to the request you are about to send.

Choose reasoning effort#

Open the reasoning menu and choose from the presets exposed for the selected model. Possible labels include No reasoning, Minimal, Low, Medium, High, Extra high, and Max, but no model is required to offer every level.

Use lower effort for direct rewriting, extraction, classification, or simple questions. Use higher effort for multi-constraint planning, difficult analysis, proofs, architecture, or tasks where checking intermediate implications matters.

Higher reasoning is not automatically more accurate. It can take longer and use more of the plan’s AI allowance. A clear prompt and relevant source material usually matter more than selecting the maximum level.

Turn on Web Search#

Select Web to request the supported provider’s hosted web search. Use it when the answer depends on current public information, such as recent events, present product details, or an up-to-date policy.

When search returns citations, the answer includes a Web Sources section with titles, domains, and links. Open the sources and verify that each one supports the nearby claim.

Request an image#

Select Image in Auto mode to route the next request to an available image model. The route treats the request as an image-generation or image-editing instruction even when the wording is conversational or not explicitly visual. For example, What is 2+2? produces a visual answer instead of being handled as an ordinary text question.

Know the boundaries#

• Web and Image are request-scoped composer routes and cannot be active together.
• Tool availability varies by model and provider.
• GPT-5 Nano does not support Web Search with its Instant preset. Choose Medium or High to make the Web control available.
• A selected tool is a request, not a promise that useful sources exist.
• Search results are external content and can be incomplete, outdated, or adversarial.
• Load context mode disables reasoning and tool generation controls because it does not call a model.

Confirm what was used#

The assistant message’s generation label records the actual model, reasoning effort, and enabled tools when available. Use that record rather than assuming the composer’s prior state. Auto mode can choose these values without showing the controls, and its completed answer exposes the resulting metadata.

Reasoning summaries are not hidden chain of thought#

The visible work log contains provider-safe progress and summary text. It is designed to explain activity without exposing private internal reasoning. Judge the answer by its result and evidence, not by the apparent length of the work log.