Skip to content

How AI reasoning models work

A reasoning model works through a problem before it answers. That measurably helps on some work, costs latency on all of it, and is frequently misunderstood. This page covers what the thinking actually is, when it is worth waiting for, and why Tartarus AI shows it to you separately.


What "thinking" actually means here

A standard model produces its answer token by token from the first word. Whatever it is going to conclude, it starts writing immediately — which means the beginning of the answer is committed before the end has been worked out.

A reasoning model is trained to generate an intermediate stretch of text first: working, considered alternatives, checks. Only then does it write the reply. The mechanism is not mysterious — it is the same next-token prediction — but it gives the model room to be wrong and correct itself somewhere that is not the answer.

That is the entire trick, and it is why it helps on problems with steps and does nothing for problems without them.

When it is worth the wait

Reasoning costs time and tokens. On a question with one obvious answer it buys nothing, and you have paid latency for the privilege.

When reasoning helps
WorkWorth it?
Multi-step problems with dependenciesYes
Comparing options against constraintsYes
Maths and quantitative workYes
Planning, architecture, structureYes
Debugging across several filesYes
Drafting and rewriting proseRarely
SummarisingNo
Quick factual questionsNo

Why the thinking is shown separately

On Tartarus AI the reasoning streams on its own channel into a collapsible panel above the reply. It is never spliced into the answer.

This matters more than it sounds. When working is mixed into the response you have to edit it out of anything you copy, and you cannot tell where the model stopped exploring and started concluding. Kept separate, the answer is clean and the working is still there when you want to check whether a conclusion was reached for a good reason.

Being able to read the working is also the fastest way to catch a confidently wrong answer, because the mistake is usually visible in the reasoning several steps before it reaches the reply.

What it does not fix

Reasoning is not verification. A model can reason fluently to a wrong conclusion, and the presence of visible working makes that conclusion more persuasive rather than more correct. Long chains can also drift, compounding an early error through every subsequent step.

It does not add knowledge either. A reasoning model that does not know something reasons carefully from not knowing it. For anything that depends on current facts, the capability you want is search.

Reasoning on Tartarus AI

Two of the three models reason. Balanced reasons briefly and is the everyday default, available on Standard and above. Deep reasons at length and is included with Pro and Ultra.

Fast does not, deliberately: it is the model for the majority of questions where an immediate good answer beats a considered one two minutes later. The model is chosen per conversation, so the decision costs a click.

Questions

Does Tartarus AI have a reasoning model?
Yes. Balanced reasons briefly and is included from Standard upwards; Deep reasons at length and is included with Pro and Ultra. Their thinking is shown in a separate collapsible panel.
What is the difference between a reasoning model and a normal one?
A reasoning model generates intermediate working before its answer, which gives it room to correct itself somewhere other than the reply. It helps on problems with steps and adds latency to everything.
Can I see the model's reasoning?
Yes, in a collapsible panel above the answer. It is streamed on its own channel and never mixed into the reply, so the text you copy is the answer alone.
Does reasoning make the answer correct?
No. A model can reason fluently to a wrong conclusion, and visible working can make a wrong answer more persuasive. It is a way to check the route, not a guarantee of the destination.
Should I always use the reasoning model?
No. On drafting, summarising and quick factual questions it costs latency for no benefit. Fast is the better default; switch to Deep when the problem has steps.