Thinking before answering
Some new models think before they answer.
They write little notes to themselves, step by step, like you do on scrap paper.
This helps a lot with puzzles and maths.
Try a puzzle. First a fast guess, then with thinking steps.
Your turn
Pick a puzzle. Press Fast, then press Think.
Toy calculation· answers written by handPip has 3 apples. Pip gets 2 more, and then eats 1. How many apples now?
These puzzle answers are written by hand to show the idea. No model is running here.
Real thinking models
- September 2024 · OpenAIo1
o1 learned to think step by step before it answered.
source: OpenAI announcement - January 2025 · DeepSeekDeepSeek-R1
DeepSeek-R1 showed that a model anyone can download could learn to reason by practising.
source: arXiv 2501.12948 - 8 April 2026 · MetaMuse Spark
Meta's Muse Spark can see what you see through a phone or smart-glasses camera. It has a quick mode and a thinking mode.
source: Meta announcement - August 2026 · AlibabaQwen3.8
Qwen3.8 has 2.4 trillion knobs and a school of 512 helpers inside. It always thinks before it answers.
source: Hugging Face model card - 3 September 2026 · OpenAIGPT-6 Astra
Some researchers worry that its thinking is harder for people to read and check.
source: TechCrunch
Thinking step by step helps with hard problems.
Grown-up corner: the precise version
- thinking notes = chain of thought (reasoning tokens)
Reasoning models are trained, mostly with reinforcement learning, to generate intermediate tokens before the final answer; spending more tokens at answer time (test-time compute) improves accuracy on hard problems.
The puzzle answers here are written by hand to show the idea; no model is running.