Thinking mode makes an AI model write out its working before it gives you an answer. That extra step makes it noticeably better at math, logic puzzles, code and plans with lots of constraints. It also makes it slower, because the model writes a lot more before the answer starts. For rewriting, chatting, translating, summarizing and simple questions, leave it off. And while the visible reasoning is handy for spotting mistakes, research shows it isn’t a reliable record of how the model actually reached its answer.
What is thinking mode in AI? #
Most chat apps now offer a switch called thinking, reasoning or something similar. ChatGPT, Gemini, Claude and DeepSeek all have one, under names that change often. Open-weight models such as Qwen 3.5 have it built in too.
A model with thinking mode has been trained to produce a scratchpad first: it restates the problem, tries an approach, checks it, sometimes backtracks, and only then writes the answer. The final answer is conditioned on all that working, so the model gets more chances to catch its own errors.
A classic test shows the difference:
A bat and a ball cost $1.10 in total. The bat costs $1.00 more than the ball. How much does the ball cost?
The quick, intuitive answer is 10 cents, and a model answering in one go often says so. With thinking on, it usually sets up the algebra (ball + ball + $1.00 = $1.10) and gets the right answer: 5 cents.
When does thinking mode help? #
| Task | Thinking mode | Why |
|---|---|---|
| Math and word problems | On | Step-by-step working catches arithmetic and setup errors |
| Logic puzzles and riddles | On | Lets the model test and reject wrong answers |
| Finding a bug in code | On | Tracing through code step by step helps |
| Plans with many constraints | On | Schedules, budgets and itineraries with rules to satisfy |
| Comparing options with trade-offs | Sometimes | Helps if there are real criteria to weigh |
| Summarizing, rewriting, changing tone | Off | Nothing to reason about; just slower |
| Translation | Off | Little benefit for the time it takes |
| Casual chat and creative writing | Off | Can make replies stiffer, not better |
| “What year did X happen?” | Off | Reasoning can’t supply a fact the model doesn’t know |
That last row matters. Thinking mode improves how a model uses what it knows. It doesn’t add knowledge. If a model doesn’t know a fact, reasoning can make it more confident in a wrong answer, not less. We explain why models invent things in why does AI make things up.
Why is thinking mode so slow? #
Because the model generates every word of its reasoning before your answer begins, and on hard problems that can be many hundreds of words.
The cost is easy to estimate. If your phone generates 15 tokens per second and the model thinks for 600 tokens, you wait about 40 seconds before the answer starts, then the answer takes its own time. A cloud model on data-center hardware hides much of that, but you still see the pause.
Long reasoning also uses part of the model’s context window, the amount of text it can keep in view, and it uses more battery on a phone because the chip works for longer. Our guide to how fast a local LLM runs on your phone explains tokens per second and how to measure it.
Can you trust the reasoning an AI shows you? #
Partly. It’s useful for checking the steps, but it isn’t a transcript of the model’s real decision process.
In an April 2025 study, Anthropic slipped hints about the answer into questions and checked whether reasoning models admitted using them. Claude 3.7 Sonnet mentioned the hint only 25% of the time on average, and DeepSeek R1 39% of the time. The researchers found models would sometimes “construct fake rationales for why the incorrect answer was in fact right.”
So read the reasoning the way you’d read a colleague’s working:
- Use it to spot errors. A wrong assumption or a dropped number is often visible in the steps.
- Don’t treat it as proof. Neat reasoning can end in a wrong answer, and a right answer can come with reasoning that isn’t how the model actually got there.
- Check anything that matters with a calculator, a source or a person. For numbers, our guide to whether local AI can do math shows how to set up problems so they’re easy to verify.
Thinking mode on your phone #
You don’t need a cloud chatbot to use reasoning. Several open models that run on phones support it.
In Personal LLM, reasoning models such as Qwen 3.5 4B and 9B have a thinking mode you can switch on per chat, or with the one-tap Thinking preset. The step-by-step working appears in a collapsible Reasoning panel above the answer, so you can read it or fold it away. Gemma 4 E4B offers hybrid thinking for harder questions. Thinking is off by default, because most questions don’t need it and replies come faster without it.
A few features make local reasoning easier to manage:
- A live tokens-per-second readout on every reply shows what the extra thinking costs on your phone.
- Switching models mid-chat lets you start on a fast model and hand a hard question to the 9B from the chat header, in the same conversation.
- Everything stays on the phone, including the reasoning, and it works offline.
The honest limit: a 4B or 9B model thinking on a phone is slower and less capable than a frontier cloud reasoning model. For everyday math, logic and code questions it’s often enough. For research-level problems, it isn’t.
How to get better results from thinking mode #
- Turn it on only for hard questions. For everything else it just adds a wait.
- Give every constraint up front. “Plan three dinners under $40 total, no nuts, 30 minutes each” gives the model something to check against.
- Ask for a clear final line. “End with the answer on its own line” keeps long reasoning from burying the result.
- Use the biggest model your phone runs comfortably. Very small models can talk themselves in circles. See how to get better answers from local AI.
- Stop it if it loops. If the reasoning repeats itself, stop the reply and rephrase the question more tightly.
Frequently asked questions #
What’s the difference between a reasoning model and a regular model? #
A reasoning model has been trained to write out step-by-step thinking before answering, which helps on multi-step problems. A regular model answers directly. Many current models, including Qwen 3.5, can do both, with thinking switched on or off.
Does thinking mode make AI more accurate? #
On math, logic, code and planning, usually yes. On facts the model doesn’t know, no, and it can make a wrong answer sound more convincing. On simple tasks like rewriting, it makes little difference.
Why does AI take so long to answer in thinking mode? #
It writes out its reasoning first, often hundreds of words, before starting the answer. Every one of those words takes time to generate, especially on a phone.
Should I leave thinking mode on all the time? #
No. Leave it off for everyday questions and turn it on when you hit a problem with several steps or a right answer you can check. That’s why many apps, Personal LLM included, keep it off by default.
Does thinking mode use more battery? #
On a phone, yes, roughly in proportion to how much longer the model runs. A long reasoning pass keeps the chip busy for longer than a short direct answer. See does local AI drain your phone battery for how to measure it.