---
title: Thinking
description: IsonAI's Thinking modes — Fast, Auto, and Deep — how adaptive effort works, the safe public thinking narrative, and how it differs from Deep Research.
last_verified: 2026-07-17
---

# Thinking

**Thinking** controls how deeply IsonAI thinks before answering. Its picker sits next to the model picker: **Fast**, **Auto**, and **Deep**.

## Fast, Auto, Deep

- **Fast** minimizes latency and answers directly, without a separate thinking stage.
- **Auto** (the default) adjusts thinking effort dynamically based on the question's complexity, risk, and calculation needs. Simple questions are answered directly; layered questions get thought through first.
- **Deep** provides a **high capability ceiling** — not a target to be spent. Questions that are already solved are not stretched out just to use up the thinking budget; what gets cut is overthinking, not quality.

## Effort as an adaptive ceiling, not a target

The effort level sets the ceiling, and IsonAI stops early once the problem is solved. Some behaviors that follow from this:

- On questions that require search, IsonAI assesses evidence sufficiency (relevance, source diversity, aspect coverage). If the initial evidence is already adequate, the answer is synthesized directly from that evidence without extra planning rounds.
- Closed-form math and analysis of material you attach do not run useless tool steps.
- Coding, connector use, research, fast-changing facts, and high-stakes topics keep the full tool loop.
- Thin or irrelevant search results still trigger further searching — economy never comes at the cost of evidence.

Each effort level also has a **first-answer deadline**. If the model is still in hidden thinking when the deadline hits, IsonAI recomposes the answer without a thinking stage — using the same context, relevant memory, and search results — so you never wait too long.

## The visible Thinking narrative

When Thinking is in use, the interface shows a **public thinking narrative** that streams phrase by phrase while the answer is composed: numbers and intermediate results in calculations, considerations and obstacles in analysis, or the failure mechanism being tested in debugging. The narrative's language follows the language of your question.

Important to understand: this narrative is a **separately produced public explanation**, not raw chain-of-thought (the model's raw internal thinking). There are four distinct layers:

1. **The model's internal thinking** — never shown.
2. **The public Thinking narrative** — the safe explanation you see streaming.
3. **The final answer** — the referenced result.
4. **Tool/search progress** — status lines while IsonAI searches or opens sources.

The narrative may use your conversation history and [memory](/en/memory/) when directly relevant, and assumptions are always **clearly labeled as assumptions**. System prompts, credentials, other users' data, and internal tool payloads are never part of the narrative. If a useful, safe narrative cannot be composed, the panel is simply not shown — not replaced with generic text.

After the answer finishes, the narrative is reconciled with the final answer, stored with the conversation, and the Thinking panel remains available in a **collapsed** state — click to open it again, including when the conversation is reopened.

## Thinking on LYT

The [LYT](/en/models/) tier always behaves as **Fast**, whatever the state of its physical backend — Auto and Deep are not available on LYT. When LYT is explicitly selected, the Thinking picker is hidden.

## Thinking Deep vs Deep Research

They are different:

- **Thinking Deep** sets the reasoning effort for **a single answer** in a normal conversation.
- **[Deep Research](/en/deep-research/)** is a separate research flow that investigates many web sources and produces a full report with references. Deep Research only activates when explicitly requested — Thinking Deep does not trigger it.

## Related

- [Models & routing](/en/models/)
- [Search & grounding](/en/search/)
- [Deep Research](/en/deep-research/)