Extended Thinking means letting Claude draw on an internal thinking resource to fully work through a problem before giving its formal reply — this is a computational resource-allocation mechanism, operating at the mechanism level. It differs from Chain of Thought, which is easy to conflate but plays a different role: chain of thought explicitly requires reasoning steps to be written into the reply itself, visible to you step by step. Extended thinking instead gives the model more room to think, but by default that thinking process isn't necessarily shown in full — what you see may just be the final answer produced after the thinking happened. Put simply, chain of thought governs whether the thinking gets written out for you to see, and extended thinking governs whether the model gets enough room to think in the first place. The two can be used together, or either one on its own.
This mechanism is needed because some tasks can still go wrong on complex, multi-step problems even when the model is given complete reasoning instructions, simply due to insufficient computational resources — scheduling problems that weigh multiple constraints at once, judgment tasks that require cross-checking several documents before reaching a conclusion, or financial calculations involving multiple layers of conditional logic. The issue in these cases isn't that the model doesn't know how to approach the problem; it's that under standard resource allocation, it doesn't think deeply or completely enough. Extended Thinking exists to give tasks like these extra thinking room, letting the model cross-check possibilities and conditions more thoroughly before producing its formal answer, reducing the chance of missing a critical detail because it moved too fast.
In practice, Extended Thinking is typically enabled through model settings or an interface option — a toggle-level decision, not something requested step by step within the prompt text. This differs from Chain of Thought, which is explicitly requested through prompt wording; extended thinking is enabled through settings or options that give the model more thinking resources to draw on. Whether to use it comes down to the nature of the task: tasks involving multiple interdependent conditions (changing one variable affects several other judgments), tasks that require cross-checking multiple sources before reaching a conclusion, or tasks that have historically shown omissions or errors under standard processing are all good candidates. By contrast, tasks with fixed formats and simple logic — filling in fields according to a template, straightforward format conversion — usually see no meaningful difference from extended thinking and just add unnecessary wait time.
For you, the real trade-off with Extended Thinking is between time cost and accuracy — enabling it usually slows the reply down, in exchange for a lower error rate on complex tasks, but whether that trade is worth it depends on how costly a mistake actually is for the task at hand. An internal draft that comes out wrong just gets edited once, a low-cost mistake not worth waiting extra time to avoid. A financial calculation or contract-clause judgment going out the door that comes out wrong can cascade into real trouble, a high-cost mistake worth the extra thinking time. What's worth noting: extended thinking doesn't show you the full thinking process by default, and if what you actually need is to verify each reasoning step yourself, what you're looking for is Chain of Thought, not extended thinking — the two get confused easily, and picking the wrong mechanism won't get you the verifiability you were after.
Anthropic's model cards and technical documentation describe how extended thinking mode gives the model more internal computational room before answering, and it has shown higher accuracy than standard mode on public benchmarks requiring multi-step reasoning, such as competition-level math problems (AIME) and graduate-level science question answering (GPQA); these benchmarks are deliberately designed around problem types that require multiple reasoning steps and are prone to mid-process errors, precisely to highlight the performance difference extended thinking makes on this category of task.
The upside is a lower error rate on complex multi-step tasks, without needing to compensate for computational limits by writing longer prompts. The downside is a longer reply time, and the thinking process isn't shown in full by default — if what you need is verifiable reasoning steps, this mechanism alone won't provide that, and for tasks with fixed formats and simple logic it brings no meaningful benefit while still adding wait time.