What is 'extended thinking' in Claude, and what types of tasks benefit most from it?

Extended thinking is a capability where Claude performs explicit internal reasoning (visible as thinking tokens in the response) before generating its final answer. This is similar to chain-of-thought but implemented at the model level with a dedicated reasoning phase. It most benefits: complex mathematics, challenging coding problems, multi-step logical reasoning, scientific analysis, and any task where working through the problem carefully before answering matters. It generally increases latency and token usage in exchange for improved accuracy on hard problems. Option A describes response length, not the reasoning process. Option C (internet search) is not what extended thinking does. Option D describes context window expansion.