Chubby♨️@kimmonismus · XAnthropic推出adaptive thinking功能，允许Claude根据请求动态分配推理token。与OpenAI通过GPT-5在不同模型间路由不同，Anthropic选择让单一模型自行调节。此举背后是企业市场利润压力与严重算力短缺——OpenAI CRO在泄露备忘录中指出Anthropic误判了计算资源采购需求。该功能虽提升效率却导致质量不稳，显示Anthropic正优先服务企业客户而牺牲消费者体验，这一点在Opus 4.7发布中已显露无遗。
A few more thoughts on Anthropic's adaptive thinking. Because it's quite revealing and offers some insight.
First of all： nobody asked for this feature. And I don't mean that as a rant. Rather， one has to ask why Anthropic implemented it directly. And the answer is， of course， as simple as it is efficient.
Profit margins aren't high enough in the consumer sector； Anthropic focuses on the enterprise and business sectors. At the same time， it's obviously bad PR when there are constant complaints that all rates are exhausted， while the competition， namely OpenAI， repeatedly increases and resets the rates.
So Anthropic wanted to do what OpenAI implemented with GPT-5： dynamic compute allocation. While OpenAI routes between different models - Instant for simple tasks， Thinking for complex ones - Anthropic's adaptive thinking lets the same model decide how many reasoning tokens are needed for the request. The idea： an efficiency gain with （ideally） consistent quality.