Thinking mode in chat: when to turn it on and when not to

Modern reasoning models — Claude, GPT, DeepSeek and others — have two ways of working: answer right away, or "think" first. In NeuralBox chat the second mode is turned on in the chat settings: before answering, the model works through the task, and its line of reasoning is shown in a separate block above the answer.

The difference isn't cosmetic: on hard tasks thinking noticeably improves the answer, while on simple ones it only slows it down and raises the price, because thinking tokens are billed the same as answer tokens. Here's when the switch is worth turning on.

How it works

With thinking mode on, the model "thinks" first: it breaks the task into steps, checks options and catches its own mistakes — all of this reasoning is visible in a separate block. Then it gives the final answer. A useful side effect: the thinking block shows how the model reached its conclusion, so it's easier to spot where it went wrong.

The mode is available for reasoning models — Claude, GPT, DeepSeek and others. It's turned on in the chat settings and can be combined with switching models right in the conversation: ask a simple question to a fast model, then switch to a flagship with thinking on for a hard one.

When to turn it on

Thinking pays off where the answer takes several connected steps and it's easy to make a mistake on the fly.

When you don't need it

For simple requests — translating a phrase, writing an email, explaining a term, rewording a paragraph — thinking isn't needed: the answer won't get better, but you'll wait longer and pay more. Thinking tokens are billed the same as the answer, so a "thinking" answer to a simple question means paying for work nobody needed.

A practical rule: keep the mode off by default and turn it on selectively, when the model gets things wrong without it or the task is clearly multi-step. Without thinking, a typical message costs about $0.01–0.05 with flagship GPT models, $0.03–0.15 with Claude Sonnet, $0.10–0.50 with Claude Opus and a fraction of a cent with DeepSeek on the Basic plan — and the longer the conversation, the more each message costs.

How to try it

Sign in with Telegram, Google or email, top up your balance (from $5 by card, from $2 in crypto) and open the Text tab. Choose a reasoning model (Claude, GPT or DeepSeek), turn on thinking mode in the chat settings and give it a task where a regular answer let you down. You can switch models mid-conversation, and the context is kept.

More about the chat, the models and prices is on the Chat with GPT, Claude and Gemini page.