Imagine asking someone for directions. Instead of answering right away, they step aside, sketch three routes on a notepad, cross two of them out — and come back with the best one. That's roughly what Thinking Models try to do.
Google has rolled out "Deep Think" for its top-tier Gemini model. It's Google's answer to OpenAI's reasoning models (now the GPT-5 series) and Anthropic's Extended Thinking: models that internally explore multiple solution paths before delivering an answer. The goal is simple — fewer mistakes on tricky tasks.
What's actually going on?
Regular AI models generate responses token by token. One word after another, no going back, no self-correction. Thinking Models get extra compute time for an "inner monologue" — a reasoning phase before the actual response. You'll often see this as a collapsible section labelled "Thinking" or "Reasoning" in the interface — the model's deliberation, made visible.
Sounds like a technical gimmick. It's not: on math problems, debugging tasks, and logical puzzles, Thinking Models consistently outperform their fast counterparts.
Think of the chess analogy: a player who thinks three moves ahead beats one who only reacts to the current position. Thinking Mode is the AI equivalent — and the industry is betting heavily on it right now.
What does it cost — and what's free?
Honest answer: full thinking power costs money.
Google's Deep Think is only for paying Ultra subscribers — currently roughly $100 to $200 a month, depending on tier. OpenAI's strongest thinking models sit behind ChatGPT Plus or Pro too. But hardly anyone needs the priciest tier — a thinking mode now ships in free tools as well.
But there are free ways to try it:
- Gemini 2.5 Flash is free at aistudio.google.com and has its own thinking mode. Not as deep as Deep Think, but noticeably better on complex questions than a standard chatbot.
- ChatGPT with o4-mini is unlocked for free users — it also thinks before answering. The free tier has limits, but enough for a first taste.
- Claude.ai occasionally offers Extended Thinking on the free plan when Anthropic activates it.
Short version: you can try this without a credit card.
Who is this for?
Thinking Mode isn't the right tool for every task. If you want to know who starred in a film, or need a quick piece of text — regular chatbot, done. No thinking required.
But if you're:
- debugging a tricky spreadsheet formula,
- working through a logic problem,
- planning something complex,
- chasing a code bug that hides on first glance,
...then Thinking Mode is worth it. It takes longer, uses more compute — and in return makes fewer mistakes.
Practical rule of thumb: if a task needs more than two reasoning steps, turn on Thinking Mode. Quick fact checks don't need it.
Does AI actually think now?
Short answer: no.
Longer answer: the model isn't running a real thought process. It's still optimizing statistically — predicting which tokens follow which — just with more intermediate steps. The "inner monologue" isn't consciousness. It's a computational detour that reduces errors.
The results are still impressive. Should we call it thinking? Philosophical question. Is it more useful than fast generation? Clearly yes.
Worth trying — and with Gemini 2.5 Flash or ChatGPT Free, it costs you nothing right now.