Extended Thinking

Claude 3.7 Sonnet features extended thinking functionality, allowing the model to engage in deeper, longer thinking before responding.

When Claude's Extended Thinking feature is enabled, the model outputs an internal reasoning process (a thinking block) before generating the final response; this reasoning is visible to the user but is not counted as part of the conversation context.

The Extended Thinking version of Gemini 3.8 Live incorporates a more thorough thinking process before voice responses, making it suitable for complex reasoning scenarios, while the standard version places greater emphasis on response speed.

Reasoning models perform internal Chain-of-Thought before returning final results, consuming large amounts of tokens and computation time in exchange for deeper understanding and lower error rates.

Excessive upfront thinking incurs increased latency and computational costs, which may not be worthwhile in scenarios that prioritize response speed.

Extended thinking mode causes the model to perform an explicit internal reasoning chain before output, at the cost of significantly increased token consumption, with reasoning tokens typically billed at a higher rate.

Claude Opus 4属于推理增强型模型，处理复杂任务时会启动扩展思维链（Extended Thinking），单次复杂推理的算力成本可达普通模型的20-50倍

The Extended Thinking feature requires passing a thinking parameter object in the API request, specifying type: 'enabled' and budget_tokens to control the depth of reasoning.

In Anthropic's Claude 3.7 and later series, the model's internal reasoning (thinking) and the final reply share the same max_tokens budget, and max_tokens is the total upper limit of 'thinking + output'.

Effort setting is an adjustment knob introduced by Anthropic in the extended thinking feature, allowing users to specify how much compute the model invests in internal reasoning before generating a response.

Extended Thinking mode increases latency, and it is officially recommended to disable this mode for simple everyday questions to achieve the fastest output

Effort tiering is a reasoning budget control mechanism introduced by Anthropic for the Claude 3.7 series, specifying the reasoning token upper limit via the budget_tokens field of the thinking parameter.

After enabling Extended Thinking in the Claude series, token consumption per request can be several times that of normal mode.

Blindly raising the thinking effort parameter does not necessarily yield better results and may instead cause resource waste and task failure.

Full chain-of-thought reasoning requires generating a large number of tokens, which is extremely demanding on computing resources.

Anthropic Claude's Extended Thinking feature explicitly requires that reasoning content must be passed back in multi-turn conversations, otherwise an error is returned.

Traditional extended thinking often comes at the cost of increased latency, creating tension with the low-latency requirements of voice scenarios

Although the reasoning tokens produced by Claude's Extended Thinking mode are not visible to users (marked as thinking_content at the API layer), they still undergo the full Decode process and are billed at the output price.

For mechanical tasks such as format conversion and simple renaming, it is recommended to directly disable thinking mode to eliminate unnecessary Decode costs.

Opus 5 enables Extended Thinking by default; thinking can only be disabled at effort levels of high or below, and the official guidance strongly recommends not disabling thinking.