Thinking

Thinking controls how IsonAI works on a single answer. Its picker sits next to the model picker: Fast, Auto, Deep, Extra Deep, and Core.

Five Thinking choices

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.
Extra Deep is available on Pro. It tests its own conclusion before accepting it: weighing an alternative interpretation, verifying a decisive figure through a second route, looking for evidence that could overturn the conclusion, and applying a higher evidence bar while searching. It is more rigorous, but usually slower than Deep.
Core is the highest-quality mode for difficult problems and is available on Pro. Core is not simply a higher effort level: it maps the goal and constraints, gathers evidence, builds several independent analyses, compares and checks them, then releases one final answer that has passed final verification. If the result cannot be verified to that standard, Core states the limitation honestly.

On PRME, Fast turns thinking off; Auto adaptively chooses between off, Low, and Medium; Deep uses native Medium; and Extra Deep uses native xhigh. Core uses native xhigh for each candidate analysis and logic-repair pass, then adds multi-stage verification. Core is therefore not simply another name for xhigh.

Core on Omni and PRME

Core runs on the explicitly selected Omni or PRME tier and never silently switches models. PRME with Core is IsonAI's strongest chat profile; Omni with Core still provides the full Core workflow on the default tier.

Core can take several minutes and has a limit of 5 uses per day for Pro accounts. It cannot currently be combined with Deep Research, image generation or editing, document editing, or image attachments. Choose another Thinking mode for those flows.

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.

The effort modes from Fast through Extra Deep also have 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. Core uses its own multi-stage workflow and shows progress until the final answer is ready.

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:

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

The narrative may use your conversation history and 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 tier always behaves as Fast — Auto, Deep, Extra Deep, and Core are not available on LYT. When LYT is explicitly selected, the Thinking picker is hidden.

Deep, Extra Deep, Core, and Deep Research

They serve different purposes:

Deep deepens one answer-composition process.
Extra Deep adds conclusion testing and a higher evidence bar to one answer.
Core runs several analyses and checks before releasing one final answer.
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; other Thinking choices do not trigger it, and Core cannot be combined with it.

Last verified: 2026-09-01