Extended Reasoning
A deliberate, operator-or-product-selected mode in which a model runs a longer internal reasoning pass before answering, trading latency and cost for depth on hard problems.
What It Is
Extended reasoning is the “think longer before you answer” mode that Google’s Deep Think embodies. When it is switched on, the model does not draft a reply on its first pass. It spends more time reasoning internally first, working through steps, checking its own intermediate conclusions, and only then producing an answer. Google put this mode behind its $250-a-month Ultra tier alongside the Gemini 3.5 Pro arrival on July 17, 2026, which is the clearest signal yet that vendors now treat depth of reasoning as a premium product, not a default.
The mechanism is more test-time compute. A model has two ways to get better at a hard problem: train a bigger model, or let the existing model spend more effort at the moment you ask. Extended reasoning is the second lever. It lengthens the model’s chain of thought, the internal sequence of steps it takes before committing to an answer, so that a genuinely hard prompt gets more deliberation than a simple one. The tradeoff is direct: you wait longer and you pay more, in exchange for a better answer on problems where the first instinct is often wrong.
How It Actually Works
The model is given room to reason across many more internal steps than a normal reply would use. On a hard math proof, a multi-file code change, or a tangled logic problem, it can explore several lines of attack, notice a dead end, back up, and try another before it ever writes the answer you see. Some implementations run several reasoning attempts in parallel and select the strongest.
That extra work is not free. Every internal step consumes tokens and compute, which is why extended reasoning is metered and gated. It is the reason a single Deep Think query can cost many times what a standard query costs, and the reason vendors put it behind their most expensive tier rather than turning it on for everyone.
The Cost and Tradeoff
The honest tradeoff is latency and money for depth. A standard reply lands in a second or two. An extended-reasoning reply can take far longer and consume a much larger token budget, so the same question can cost pennies or dollars depending on which mode answers it. That is fine when the problem is genuinely hard and being wrong is expensive. It is waste when the problem was simple and a fast pass would have been correct.
The failure mode is using the expensive mode as a security blanket. Turning on extended reasoning for a task that never needed it buys you a slower, pricier version of an answer you already would have gotten.
Common Misconceptions
The most common confusion is between extended reasoning and adaptive thinking. They are opposites in one crucial way. Adaptive thinking is the model deciding on its own how hard to think, scaling its effort up or down without anyone touching a switch. Extended reasoning is a mode that is deliberately turned on, by the operator or by the product tier, and stays on for that call regardless of whether the prompt needed it. One is automatic self-pacing; the other is a lever a human or a paywall pulls.
The second misconception is that longer reasoning always means a better answer. It does not. Past a point, more internal steps add cost without adding correctness, and on easy prompts they can even talk the model out of a right first answer.
How TWO Uses It
TWO’s line on extended reasoning is that it is a tool for a specific class of problem, not a default posture. The operator question is never “do I want the smartest mode,” because the honest answer is always yes and the honest bill is always higher. The question is “is this problem hard enough that a wrong first answer would cost me more than the extended-reasoning premium.” If the answer is no, the fast pass wins.
Scott’s Take: Pay for the model to think longer only when being wrong would cost you more than the extra minute and the extra dollar, and never a cent before.
In practice, Scott’s rule is to draft with the standard mode and reserve extended reasoning for the two or three genuinely hard calls in a project: the architecture decision, the proof that has to hold, the analysis where a subtle error propagates everywhere. Everything else runs on the cheap, fast pass. The discipline is treating the expensive mode as a scalpel, not a mood.
What to Watch Next
Watch whether extended reasoning stays a paid tier or collapses back into the default as compute gets cheaper. When a lab can afford to think longer for everyone, the $250 gate becomes a marketing line rather than a cost floor. Watch, too, whether vendors start showing you the reasoning they charged you for. A frontier model that lets you audit the longer pass is selling a tool you can trust; one that hides it and only shows the bill is selling a black box at a premium.