Reasoning · February 25, 2025 · 2 min read

Extended Thinking: Inside Claude 3.7 Sonnet

Anthropic's Claude 3.7 Sonnet introduced a hybrid approach: it can answer quickly for everyday requests or engage extended thinking for harder problems, with developers able to control how much reasoning to spend.

One model, two modes

Combining quick and deliberate reasoning in a single model simplifies architecture. Instead of routing between separate models, teams can dial reasoning up or down per request based on difficulty and budget.

Practical evaluation still wins

More reasoning isn't always better — it costs time and money. The teams that benefit most measure where extended thinking actually improves outcomes and apply it selectively.

Our take: controllable reasoning is a powerful lever. We build evaluation into client systems so that lever is tuned with data, not guesswork.