<p>Anthropic's extended thinking feature, which gives Claude additional tokens to reason through problems before producing a final response, has been deployed long enough for a meaningful empirical picture to emerge. The verdict is nuanced: significant gains on specific task types, minimal benefit on others, and real cost implications.</p>
<h2>Where Extended Thinking Helps</h2>
<p>The clearest wins are on multi-step logical reasoning, complex mathematics, and planning tasks that require holding many constraints simultaneously. On competition-level math problems, extended thinking raises Claude's solve rate from approximately 45% to 72%. On complex coding tasks requiring architectural thinking across multiple files, success rates improve by 25-35%.</p>
<p>Legal analysis, scientific literature synthesis, and strategic planning — tasks where a human expert would explicitly reason through sub-questions before answering — also show consistent improvement with extended thinking enabled.</p>
<h2>Where It Doesn't Help</h2>
<p>For factual questions, summarization, translation, writing assistance, and most conversational tasks, extended thinking adds latency and cost without measurable quality improvement. The baseline model handles these well without additional reasoning steps.</p>
<h2>Cost and Latency Trade-offs</h2>
<p>Extended thinking consumes 3-8x more tokens than standard inference, translating to proportionally higher API costs. Response latency increases from 1-3 seconds to 10-30 seconds depending on problem complexity. For most production applications where users expect fast responses, this is prohibitive.</p>
<p>The practical recommendation from developers who've tested both modes: use extended thinking for asynchronous, high-stakes tasks where accuracy matters more than speed. Use standard inference for real-time user-facing applications. The distinction maps roughly onto human analogues: slow, careful thinking for hard problems; fast, intuitive responses for routine ones.</p>