Anthropic Economic Index report: Learning curvesAnthropic 经济指数报告：学习曲线

The Anthropic Economic Index uses our privacy-preserving data analysis system to track how Claude is being used across the economy. It’s part of our effort to understand the economic impacts of AI as early as possible, so that researchers and policymakers have adequate time to prepare.Anthropic 经济指数利用我们的隐私保护数据分析系统，追踪 Claude 在整个经济领域的使用情况。这是我们尽早了解人工智能经济影响的努力之一，旨在为研究人员和政策制定者留出充足的准备时间。

This latest report studies Claude usage in February 2026, building on the economic primitives framework introduced in our previous report (which used data from November 2025). Our sample covers February 5 to February 12, three months following the release of Claude Opus 4.5 and coincident with the release of Claude Opus 4.6.这份最新报告基于我们上一份报告（使用 2025 年 11 月的数据）中引入的经济基元框架，研究了 2026 年 2 月 Claude 的使用情况。我们的样本覆盖了 2 月 5 日至 2 月 12 日，即 Claude Opus 4.5 发布三个月后，且与 Claude Opus 4.6 的发布时间重合。

We first document how usage has changed relative to our previous reports: the rate of augmentation, collaborative interaction where the AI complements the user’s abilities, increased slightly in both Claude.ai and API traffic. In Claude.ai, usage diversified, with the top 10 tasks accounting for a smaller share of usage last month than in November 2025. As a result of this diversification, the average conversation in Claude.ai had a slightly lower-wage task than in previous reports.我们首先记录了使用情况相较于之前报告的变化：在 Claude.ai 和 API 流量中，增强（即人工智能辅助用户能力的协作互动）的比例略有上升。在 Claude.ai 中，使用场景更加多元化，上个月前 10 大任务所占的流量比例较 2025 年 11 月有所下降。由于这种多元化，Claude.ai 的平均对话任务工资水平略低于之前的报告。

We then focus on an important determinant of Claude’s impact on the labor market and the broader economy: learning curves in Claude adoption. We present evidence that high-tenure users have developed habits and strategies that allow them to better harness Claude’s capabilities. Indeed, we document that more experienced users not only attempt higher-value tasks, but are also more likely to elicit successful responses in their conversations.随后，我们重点关注了影响 Claude 对劳动力市场和更广泛经济影响的一个重要决定因素：Claude 使用过程中的学习曲线。我们提供的证据表明，资深用户已经形成了能够更好地发挥 Claude 能力的习惯和策略。事实上，我们记录到，经验更丰富的用户不仅尝试更高价值的任务，而且在对话中获得成功反馈的可能性也更高。

What has changed since our last report自上一份报告以来发生了哪些变化

In the first chapter, we revisit findings from our previous Economic Index report, published in January 2026. We find that:在第一章中，我们回顾了 2026 年 1 月发布的上一份经济指数报告中的发现。我们发现：

• Use cases on Claude.ai diversified. Coding tasks continue to migrate from augmentative usage in Claude.ai to more automated workflows in our first-party API traffic.¹ In this report, Claude.ai usage was less concentrated: the top 10 tasks made up 19% of all traffic in February, down from 24% in November. That said, almost all tasks in this sample appeared in at least one of our previous samples. About 49% of jobs have seen at least a quarter of their tasks performed using Claude.Claude.ai 上的使用场景趋于多元化。编码任务正从 Claude.ai 的增强型使用迁移到我们第一方 API 的自动化工作流中。1 在本报告中，Claude.ai 的使用集中度有所下降：2 月份前 10 大任务占总流量的 19%，低于 11 月的 24%。尽管如此，该样本中几乎所有任务都出现在我们之前的样本中。约 49% 的职业中，至少有四分之一的任务是通过 Claude 完成的。
• Claude adoption broadened to lower-wage tasks. As use cases have diversified, the average economic value of work done on Claude—as measured by US wages paid to workers in the associated occupations—has decreased slightly. This is caused, mechanically, by a rise in personal queries around sports, product comparisons, and home maintenance. The pattern is consistent with a standard “adoption curve” story, in which early-adopters favor specific high-value uses like coding, and later adopters take on a much wider range of tasks.Claude 的应用范围扩展到了低薪任务。随着使用场景的多元化，Claude 上所完成工作的平均经济价值（以相关职业的美国工资水平衡量）略有下降。从机制上讲，这是因为围绕体育、产品比较和家庭维护等个人查询的增加。这一模式符合标准的“采用曲线”理论，即早期采用者倾向于编码等特定的高价值用途，而后期采用者则承担更广泛的任务。
• Inequality in global usage has persisted. Usage remains heavily concentrated: the top 20 countries account for 48% of all per-capita usage, up from 45%, underscoring a persistent gap in global adoption. However, Claude usage per capita continued to converge within the United States: the share of usage accounted for by the 10 highest usage states decreased from 40% to 38% since our last report.全球使用不平等现象持续存在。使用情况仍然高度集中：前 20 个国家占人均总使用量的 48%（之前为 45%），凸显了全球应用差距的持续存在。然而，美国国内的人均 Claude 使用量持续趋于均衡：自上一份报告以来，使用量最高的 10 个州所占的份额已从 40% 降至 38%。

Learning curves学习曲线

A central finding in the Economic Index is that early adoption of Claude is very uneven: Claude is used more intensely in high-income countries, within the US in places with more knowledge workers, and for a relatively small set of specialized tasks and occupations.经济指数的一个核心发现是，Claude 的早期采用非常不均衡：Claude 在高收入国家、美国境内知识工作者较多的地区，以及针对少数特定专业任务和职业的使用更为密集。

An important question is how inequality of adoption might determine where and to whom the benefits of AI will accrue. If, for example, effective AI use requires complementary skills and expertise—which we argued in our previous report—and if such skills can be acquired through use and experimentation, then the benefits from early adoption may be self-reinforcing.一个重要的问题是，采用的不平等性如何决定人工智能的收益流向何处以及流向谁。例如，正如我们在上一份报告中所述，如果有效使用人工智能需要互补的技能和专业知识，并且这些技能可以通过使用和实验获得，那么早期采用带来的收益可能会自我强化。

In our second chapter we investigate how users appear to shape the value that they get out of Claude: how they match model capability to the task at hand, and how usage patterns and outcomes shift with experience on the platform.在第二章中，我们研究了用户如何塑造他们从 Claude 中获得的价值：他们如何将模型能力与手头任务相匹配，以及使用模式和结果如何随着平台经验的积累而发生变化。

• Model selection matches the task. We show that users choose our most intelligent model class, Opus, for tasks that normally receive higher wages in the labor market. For example, among paying Claude.ai users, Opus is used 4 percentage points more than average for coding tasks and 7 percentage points less than average for tutoring-related tasks. This model switching is about twice as stark for API users. 模型选择与任务匹配。我们发现，用户在劳动力市场中通常获得较高工资的任务上，会选择我们最智能的模型系列 Opus。例如，在付费的 Claude.ai 用户中，Opus 在编码任务中的使用率比平均水平高出 4 个百分点，而在辅导类任务中则比平均水平低 7 个百分点。对于 API 用户，这种模型切换现象更为明显，程度大约是前者的两倍。
• Higher tenure, higher success. In general, the most seasoned Claude users employ it more often for higher-education tasks and less often for personal use cases. For example, people who have been using Claude for 6 months or more have 10% fewer personal conversations and a 6% higher education level reflected in their inputs. Most strikingly, people in this higher-tenure group have a 10% higher success rate in their conversations, an association that is not explained by their task selection, country of origin, or other factors. While this could reflect sophistication of early adopters, it could also be evidence of learning-by-doing, where people get better at using Claude through experience.资历越深，成功率越高。总体而言，最资深的 Claude 用户更多地将其用于高等教育任务，而较少用于个人用途。例如，使用 Claude 6 个月或更长时间的用户，个人对话减少了 10%，输入中反映的教育水平提高了 6%。最引人注目的是，这一资深群体在对话中的成功率高出 10%，这种关联性无法用任务选择、原籍国或其他因素来解释。虽然这可能反映了早期采用者的成熟度，但也可能是“干中学”的证据，即人们通过经验提高了使用 Claude 的能力。

What has changed since our last report

Diversification of use cases in Claude.aiClaude.ai 使用场景的多元化

We first look at the kinds of tasks that Claude is asked to perform. We use our privacy-preserving system, which allows us to describe behavior at an aggregated level without revealing the content of individual transcripts. We sample 1 million conversations from both Claude.ai, our consumer-facing web product, and our first-party API, the developer-facing interface for integrating Claude into products and workflows.²我们首先考察了 Claude 被要求执行的任务类型。我们使用隐私保护系统，在不泄露个人对话内容的情况下描述聚合层面的行为。我们从面向消费者的网页产品 Claude.ai 和面向开发者的第一方 API（用于将 Claude 集成到产品和工作流中）中各抽取了 100 万条对话样本。2

Coding remains the most common use on our platforms, with tasks associated with Computer and Mathematical occupations accounting for 35% of conversations on Claude.ai (see Appendix).³ However, between November 2025 and February 2026, use cases on Claude.ai became less concentrated: the top 10 most common O*NET tasks went from 24% of conversations to just 19% (Figure 1.1).编码仍然是我们平台最常见的用途，与计算机和数学职业相关的任务占 Claude.ai 对话的 35%（见附录）。3 然而，在 2025 年 11 月至 2026 年 2 月期间，Claude.ai 上的使用场景集中度有所下降：前 10 大最常见的 O*NET 任务在对话中的占比从 24% 降至 19%（图 1.1）。

This decline in concentration partly reflects coding tasks migrating from Claude.ai to our first-party API, where Claude Code has grown to represent a large share of sampled traffic. Claude Code’s agentic architecture splits coding work into smaller API calls, which are labeled as distinct tasks. So while coding’s overall share of API traffic has grown, it is spread across many task categories rather than concentrated in a few. As a result, task concentration in the API remained roughly flat despite the influx of coding activity.这种集中度的下降部分反映了编码任务从 Claude.ai 向我们第一方 API 的迁移，Claude Code 在其中已占据了采样流量的很大份额。Claude Code 的智能体架构将编码工作拆分为更小的 API 调用，这些调用被标记为不同的任务。因此，虽然编码在 API 流量中的总体份额有所增长，但它分散在许多任务类别中，而不是集中在少数几个类别上。结果，尽管编码活动大量涌入，API 中的任务集中度仍基本保持平稳。

This migration of code out of Claude.ai is not the only factor driving decreased concentration. Part of the drop is due to changes in the mix of use cases between the two periods. Coursework fell from 19% to 12% of conversations, while personal use rose from 35% to 42% of conversations. Some of the drop in coursework can be explained by academic calendars in countries where students were on winter break during our sample period.⁴ At the same time, increasing signups beginning around February brought more casual AI users.代码从 Claude.ai 的迁移并不是导致集中度下降的唯一因素。部分下降归因于两个时期之间使用场景组合的变化。课程作业在对话中的比例从 19% 降至 12%，而个人用途从 35% 升至 42%。课程作业比例下降的部分原因可以解释为我们采样期间，部分国家的学生正处于寒假。4 与此同时，自 2 月份开始增加的注册量带来了更多休闲型人工智能用户。

While the spread of Claude's work tasks became more diverse, almost all of these had been seen before in our data. In our previous report, we noted that 49% of jobs had seen at least a quarter of their tasks performed using Claude. In this data pull, that cumulative estimate barely changed (Appendix Figure A.2). Our data from this report showed many fewer novel O*NET tasks than in our previous report.虽然 Claude 的工作任务分布变得更加多样化，但几乎所有这些任务在我们的数据中之前都出现过。在上一份报告中，我们指出 49% 的职业中至少有四分之一的任务是通过 Claude 完成的。在本次数据提取中，这一累计估算值几乎没有变化（附录图 A.2）。本报告的数据显示，新型 O*NET 任务比上一份报告要少得多。

Since our first report, we have classified conversations into one of five interaction types—directive, feedback loop, task iteration, validation, and learning—which we group into two broader categories: automation and augmentation.⁵ Figure 1.3 shows that augmentation in Claude.ai increased slightly. This was driven by small bumps in validation and learning patterns. In Appendix Figure A.3, we show that automation decreased sharply in the 1P API data.自第一份报告以来，我们将对话分为五种交互类型——指令、反馈循环、任务迭代、验证和学习，并将它们归纳为两个更广泛的类别：自动化和增强。5 图 1.3 显示 Claude.ai 中的增强功能略有增加。这是由验证和学习模式的小幅增长所驱动的。在附录图 A.3 中，我们显示 1P API 数据中的自动化程度大幅下降。

Our API platform continued to gain a relatively higher share of Computer and Mathematical tasks (usage shares by job categories are shown in the Appendix). Since August 2025, the share of tasks in this category has increased by 14% in the API and decreased by 18% in Claude.ai. As we note in our report on labor market impacts, we expect that this migration from Claude.ai to the API may signal more imminent transformation of work for the associated jobs. The increase in tasks associated with Management occupations in Claude.ai, which went from 3 to 5% of its traffic, comes from a mix of both analytical tasks (e.g., preparing an investment memo) and responding to customer questions.我们的 API 平台继续在计算机和数学任务中获得相对较高的份额（各职业类别的使用份额见附录）。自 2025 年 8 月以来，该类别的任务份额在 API 中增加了 14%，而在 Claude.ai 中减少了 18%。正如我们在劳动力市场影响报告中所指出的，我们预计这种从 Claude.ai 到 API 的迁移可能预示着相关职业工作方式即将发生更紧迫的变革。Claude.ai 中与管理职业相关的任务增加（从流量的 3% 升至 5%），来自于分析任务（例如准备投资备忘录）和回复客户问题两者的结合。

Another way to measure the change in the mix of tasks done on Claude is to look at the change in the average value of tasks, which we define as the average hourly wage of US workers who perform that task (Figure 1.4).⁶ This estimate of the value of tasks in Claude.ai has dropped slightly from $49.3 to $47.9 mostly due to an increase in simple factual questions (e.g., sports outcomes, weather) and a decrease in coding as it shifts to the API. As mentioned in our previous report, the tasks we see on Claude tend to require higher education. The plot shows that these tasks also tend to earn higher wages than the US national average.衡量 Claude 上任务组合变化的另一种方法是观察任务平均价值的变化，我们将其定义为执行该任务的美国工人的平均时薪（图 1.4）。6 Claude.ai 上的任务价值估算值从 49.3 美元小幅降至 47.9 美元，这主要是由于简单的实事查询（如体育比赛结果、天气）增加，以及编码任务随着向 API 迁移而减少。正如我们在上一份报告中所述，我们在 Claude 上看到的任务往往需要更高的教育水平。图表显示，这些任务的收入也往往高于美国全国平均水平。

While slight, changes in several primitives between the previous and current report capture similar declines in task complexity on Claude.ai. The average years of education required for the human inputs declined from 12.2 to 11.9 years, users granted more autonomy to the AI, and the time required for the human to do the task alone fell by about 2 minutes. One change goes ostensibly in the opposite direction: the tasks performed by Claude were judged to be slightly less possible for a human without access to AI.虽然变化很小，但上一份报告与当前报告之间几个基元的变化捕捉到了 Claude.ai 上任务复杂性的类似下降。人类输入所需的平均教育年限从 12.2 年降至 11.9 年，用户赋予人工智能的自主权更多，而人类独自完成任务所需的时间减少了约 2 分钟。有一个变化表面上朝着相反的方向发展：Claude 执行的任务被认为对于没有人工智能辅助的人类来说，完成的可能性略有降低。

Emergent automation patterns新兴的自动化模式

As tasks migrate to the API, they may become more exposed to automation. API workflows are far more likely to be directive, with less need for a human in the loop. In a previous report, we highlighted that customer service tasks, including, for example, automated support for payment and billing issues, are prevalent in the API data. These contributed to a higher observed exposure for Customer Service Representatives—Claude was recorded doing a high share of their tasks in automated workflows, so these jobs may be more likely to change as AI diffuses.随着任务迁移到 API，它们可能更容易受到自动化的影响。API 工作流更有可能是指令性的，对人工介入的需求较少。在上一份报告中，我们强调了客户服务任务（例如针对支付和账单问题的自动支持）在 API 数据中非常普遍。这导致客户服务代表的暴露度更高——Claude 被记录在自动化工作流中完成了他们很大一部分的任务，因此随着人工智能的普及，这些工作岗位可能更容易发生变化。

We highlight two API workflows that appeared more frequently in February as compared to three months prior, with their shares at least doubling in our latest sample:⁷我们重点介绍了两个在 2 月份比三个月前出现频率更高、在最新样本中份额至少翻倍的 API 工作流：7

• Business sales & outreach automation: sales enablement generation, B2B lead qualification research, customer data enrichment, cold-email drafting.业务销售与外联自动化：销售赋能生成、B2B 线索资格研究、客户数据丰富、冷邮件起草。
• Automated trading & market ops: monitor markets or positions, propose specific investments, inform traders of market conditions, and related tasks.自动交易与市场运营：监控市场或头寸、提出特定投资建议、向交易员告知市场状况及相关任务。

Revisiting geographic convergence重温地理趋同

In our previous report, we noted that the Anthropic AI Usage Index (AUI), which adjusts usage by a geography’s working-age population, was converging rapidly across US states: states with initially lower usage per capita showed faster adoption.在上一份报告中，我们指出根据地理区域劳动年龄人口调整使用量的 Anthropic AI 使用指数 (AUI) 在美国各州之间正迅速趋同：最初人均使用量较低的州显示出更快的采用速度。

The left panel of Figure 1.6 shows that this convergence continued in our most recent data, but at a slower pace. From August 2025 to February 2026, the share of per-person usage going to the top five states has decreased from 30 to 24%. The Gini coefficient has fallen since August 2025, though the pace of convergence has slowed. When we update our estimates from the previous report, we find that at this rate states would arrive at roughly equal usage per capita in 5–9 years, rather than 2–5.⁸图 1.6 的左侧面板显示，这种趋同在我们的最新数据中仍在继续，但步伐放缓。从 2025 年 8 月到 2026 年 2 月，前五个州的人均使用份额从 30% 降至 24%。基尼系数自 2025 年 8 月以来有所下降，尽管趋同速度有所减缓。当我们更新上一份报告的估算值时，发现按此速度，各州在 5-9 年内而非 2-5 年内将达到大致相等的人均使用量。8

Across countries (right panel), the pattern is reversed: usage has become slightly more concentrated, with the Gini rising over the same period. The countries using Claude the most (per capita) now account for a larger share of overall usage, with the top 20 countries going from 45 to 48% of usage adjusted for population.在各国之间（右侧面板），模式正好相反：使用情况变得略微更加集中，基尼系数在同一时期有所上升。Claude 使用量最多（人均）的国家现在占总使用量的份额更大，前 20 个国家占人口调整后使用量的份额从 45% 升至 48%。

Learning to use AI学习使用人工智能

In this chapter, we focus on two features of usage that reflect how people deploy and learn to work with AI: model selection and the habits of longtime users.在本章中，我们关注反映人们如何部署和学习使用人工智能的两个使用特征：模型选择和长期用户的习惯。

First, we shed light on the demand for intelligence by studying when people choose Opus, our most performant model class. Little is known about how AI users choose between different models, navigating tradeoffs around speed, performance, and cost. If users are calibrating to the task at hand, we should see Opus concentrated on harder, higher-value work.首先，我们通过研究人们何时选择我们性能最强的模型系列 Opus 来阐明对智能的需求。关于人工智能用户如何在不同模型之间进行选择，以及如何在速度、性能和成本之间进行权衡，目前知之甚少。如果用户在根据手头的任务进行调整，我们应该会看到 Opus 集中在更困难、更高价值的工作上。

Next, we study how usage differs according to tenure, finding differences across users who signed up at different times. This sheds light on learning curves: do experienced users get better over time? How does their usage differ? We find evidence consistent with learning-by-doing. Not only do higher-tenure users have greater success in their conversations, they also collaborate with Claude more, bring more challenging tasks to Claude, and are more likely to use Claude for work purposes and for a wider range of tasks.接下来，我们研究使用情况如何根据资历而有所不同，并发现了不同时间注册的用户之间的差异。这揭示了学习曲线：经验丰富的用户会随着时间的推移变得更好吗？他们的使用方式有何不同？我们发现了与“干中学”一致的证据。资深用户不仅在对话中获得了更大的成功，而且与 Claude 的协作更多，带给 Claude 的任务更具挑战性，并且更有可能将 Claude 用于工作目的和更广泛的任务。

Model selection模型选择

The different Claude model classes (Haiku, Sonnet, and Opus) offer tradeoffs in terms of cost, speed, and performance. The Opus class of models uses the most tokens and excels at complex tasks, but at a higher per-token price on our API. If users are aware of this and mindful of costs and usage limits, they should bring their most complicated and valuable tasks to Opus, while selecting other models for simpler tasks. This is broadly what we observe in the data.不同的 Claude 模型系列（Haiku、Sonnet 和 Opus）在成本、速度和性能方面提供了不同的权衡。Opus 系列模型使用的 Token 最多，擅长复杂任务，但在我们的 API 上单价更高。如果用户意识到这一点并注意成本和使用限制，他们应该将最复杂和最有价值的任务交给 Opus，同时为更简单的任务选择其他模型。这大体上就是我们在数据中观察到的情况。

Figure 2.1 below shows that, for paid Claude.ai accounts, which have access to all model classes, 55% of Computer and Mathematical tasks (like coding software) use Opus, compared to 45% of Educational tasks. Technical users may notice performance gains and actively switch away from Sonnet, the default. Or efficiency-minded users may learn to use Sonnet for simpler tasks to avoid hitting usage limits. Relatedly, the differences here could reflect that most educational tasks are already fairly easy for Sonnet, or that students are more likely to be mindful of usage limits.下方的图 2.1 显示，对于拥有所有模型系列访问权限的付费 Claude.ai 账户，55% 的计算机和数学任务（如编写软件）使用 Opus，而教育任务中这一比例为 45%。技术型用户可能会注意到性能提升并主动从默认的 Sonnet 切换。或者，注重效率的用户可能会学会使用 Sonnet 处理简单任务，以避免达到使用限制。相关地，这里的差异可能反映出大多数教育任务对 Sonnet 来说已经相当简单，或者学生更可能注意使用限制。

Figure 2.2 below shows this in a more granular way. When users perform tasks associated with higher-paid jobs, they use Opus more often. For example, on Claude.ai, 34% of Software Developer tasks involve Opus compared to just 12% of Tutor tasks. Overall, for every additional $10 of hourly wage for a task, the share of conversations using Opus increases by 1.5 percentage points for Claude.ai users. The 1P API traffic shows much more response to the complexity of the task. Its slope is about twice as large, with the Opus share increasing 2.8 percentage points for every $10 in task value. Users deploying programmatic workflows may have more reason to switch between models compared to web users.下方的图 2.2 以更细致的方式展示了这一点。当用户执行与高薪工作相关的任务时，他们更频繁地使用 Opus。例如，在 Claude.ai 上，34% 的软件开发人员任务涉及 Opus，而导师任务仅为 12%。总体而言，任务时薪每增加 10 美元，Claude.ai 用户使用 Opus 的对话份额就会增加 1.5 个百分点。1P API 流量对任务复杂性的反应要强烈得多。其斜率大约大两倍，任务价值每增加 10 美元，Opus 份额增加 2.8 个百分点。与网页用户相比，部署程序化工作流的用户可能有更多理由在模型之间进行切换。

Learning curves

The first Claude model was released in March 2023. Since then, the userbase on both Claude.ai and the API has grown rapidly. Our latest sample contains a mix of users, some of whom signed up for the very first Claude release, and some of whom signed up just the day before we measured their usage. How does one’s tenure with Claude shape their experience with it?⁹第一个 Claude 模型于 2023 年 3 月发布。从那时起，Claude.ai 和 API 上的用户群迅速增长。我们的最新样本包含各种用户，其中一些人是在 Claude 首次发布时注册的，而另一些人则是在我们测量其使用情况的前一天才注册的。一个人使用 Claude 的资历如何塑造他们的体验？9

Table 2.1 shows differences between low tenure and high tenure users, where the latter group is defined as having signed up for Claude at least 6 months ago and the low tenure users are everyone else.¹⁰ High tenure users are more likely to use Claude to iterate on their work, and much less likely to delegate greater responsibility through directive use patterns. They are 7 percentage points more likely to be using Claude for work, and use Claude for tasks that tend to require higher levels of education. Finally, their usage is less concentrated in certain tasks. The top 10 O*NET tasks account for a slightly lower (20.7% compared to 22.2%) share of usage for the high tenure group.表 2.1 显示了低资历和高资历用户之间的差异，其中后者定义为至少在 6 个月前注册 Claude，而低资历用户则是其他人。10 高资历用户更有可能使用 Claude 来迭代他们的工作，而通过指令模式委派更大责任的可能性则要小得多。他们将 Claude 用于工作的可能性高出 7 个百分点，并且倾向于将 Claude 用于需要更高教育水平的任务。最后，他们的使用集中在某些任务上的程度较低。前 10 大 O*NET 任务在高资历组的使用份额略低（20.7% 对比 22.2%）。

Below, we dig more into two of the primitives discussed above: the human years of schooling associated with each conversation, and the share of transcripts devoted to personal use.下面，我们进一步深入探讨上述讨论的两个基元：每次对话相关的人类受教育年限，以及用于个人用途的对话份额。

In the panel on the left, we show that the years of schooling needed to understand the human prompt increases by almost 1 year for every additional year of Claude usage. In the panel on the right, we show that at the same time, personal use decreases: people who signed up a year ago devote 38% of their conversations to personal use cases, compared to 44% for the newest users.在左侧面板中，我们展示了理解人类提示所需的受教育年限随着 Claude 使用年限的增加而增加，每增加一年使用量，受教育年限增加近 1 年。在右侧面板中，我们展示了与此同时，个人用途有所减少：一年前注册的人将 38% 的对话用于个人用途，而最新用户则为 44%。

Several factors could account for these patterns in the user base of a rapidly advancing all-purpose technology. The high-tenure users are self-selected and the differences here could reflect stable characteristics. They may be computer programmers, for example, who were more likely to be early adopters. Further, there’s an inherent survivorship bias: people who signed up a year before our data pull may be seeing positive results from their usage. We do not observe people who signed up a year ago but are no longer using Claude. 有几个因素可以解释这种快速发展的通用技术用户群中的这些模式。高资历用户是自选的，这里的差异可能反映了稳定的特征。例如，他们可能是计算机程序员，更有可能是早期采用者。此外，存在固有的幸存者偏差：在我们提取数据前一年注册的人可能从使用中看到了积极的结果。我们没有观察到一年前注册但不再使用 Claude 的人。

The findings mirror what we saw in our Economic Primitives report: lower-income, less educated countries paradoxically showing more complex use in some cases. The earliest adopters often have high-value, technical use cases. In poorer countries with much lower adoption, these early adopters still dominate the user base.这些发现反映了我们在《经济基元》报告中看到的情况：收入较低、教育程度较低的国家在某些情况下反而表现出更复杂的使用情况。最早的采用者通常有高价值的技术性用例。在采用率低得多的较贫穷国家，这些早期采用者仍然占据用户群的主导地位。

More casual usage emerges when AI has diffused to a broader share of the population. Indeed, among request clusters, tasks with highest mean tenure included: AI research, git operations, revising manuscripts, and startup fundraising. The tasks with the lowest average tenure have more simple workflows like writing haikus, checking sports scores, and suggesting food for a party.¹¹当人工智能普及到更广泛的人群时，更休闲的使用方式就会出现。事实上，在请求聚类中，平均资历最高的任务包括：人工智能研究、git 操作、修订手稿和初创公司融资。平均资历最低的任务则包括编写俳句、查看体育比分和为聚会建议食物等简单工作流。11

Experience effects经验效应

We explore these relationships more in Figure 2.4 below, using the log-level data to control granularly for features of the conversation. In the top panel, specification (1) shows a simple bivariate regression with task success as the outcome and the long-tenure indicator as the predictor. Success is Claude’s assessment of whether the conversation was successful, described in our previous report. The plot shows that long-tenure users are about 5 percentage points more likely to have a successful conversation.我们在下方的图 2.4 中进一步探讨了这些关系，使用对数级数据对对话特征进行细致控制。在顶部面板中，规范 (1) 显示了一个简单的双变量回归，以任务成功率为结果，以长期资历指标为预测变量。成功率是 Claude 对对话是否成功的评估，我们在上一份报告中对此进行了描述。图表显示，长期资历用户的对话成功率高出约 5 个百分点。

This could reflect that higher-tenure users are better at prompting. But what if it reflects that they bring different tasks to Claude—ones more likely to be successful?这可能反映了高资历用户更擅长提示。但如果这反映了他们带给 Claude 不同的任务——那些更容易成功的任务呢？

In Specification (2), we include fixed effects for specific O*NET tasks and request clusters. This amounts to comparing high- and low-tenure users within the same narrowly defined task, rather than across tasks. For instance, we have a request cluster called “Perform corporate financial analysis, valuation, and modeling for specific companies.” The fixed effects compare high- and low-tenure users within that cluster, and likewise within every other cluster. We would only observe a positive coefficient if, on average, long-tenure users are more successful in these within-task comparisons. This control moderates the effect somewhat, bringing it closer to 3 percentage points.在规范 (2) 中，我们纳入了特定 O*NET 任务和请求聚类的固定效应。这相当于在相同定义狭窄的任务内比较高资历和低资历用户，而不是跨任务进行比较。例如，我们有一个名为“为特定公司执行企业财务分析、估值和建模”的请求聚类。固定效应在每个聚类内比较高资历和低资历用户。只有当长期资历用户在这些任务内比较中平均更成功时，我们才会观察到正系数。这种控制在一定程度上缓和了影响，使其更接近 3 个百分点。

Finally, we ask whether this relationship is affected by higher-tenure users selecting different models, communicating in different languages, having different use cases, or signing on in different countries. This regression yields a slightly higher impact of high tenure, suggesting a 4-percentage-point higher success rate accounting for the full controls.最后，我们询问这种关系是否受到高资历用户选择不同模型、使用不同语言交流、有不同用例或在不同国家登录的影响。该回归得出了高资历略高的影响，表明在考虑完全控制的情况下，成功率高出 4 个百分点。

These results suggest that high-tenure users have more success in their Claude conversations, and that this is not due to simple factors like language or the task being performed. One intriguing potential explanation is that these users have better learned to extract what they want from AI. Facility with these platforms may be a key determinant of success that appears to scale with experience.这些结果表明，高资历用户在 Claude 对话中更成功，而且这并不是由于语言或执行任务等简单因素造成的。一个有趣的潜在解释是，这些用户已经更好地学会了如何从人工智能中提取他们想要的东西。对这些平台的熟练程度可能是随着经验积累而扩展的关键成功决定因素。

Discussion讨论

This report revisited the core measures we use to track Claude usage, and analyzed model selection and success for the first time. Since August 2025, 1P API usage has become more concentrated, with the top 10 O*NET tasks now accounting for 33% of traffic, up from 28%. Claude.ai tasks, on the other hand, have diversified since our November 2025 data. Faster adoption among low-usage states continued in the US, though at a slower pace than in the previous report. Low-adoption countries fell slightly further behind.本报告回顾了我们用于追踪 Claude 使用情况的核心指标，并首次分析了模型选择和成功率。自 2025 年 8 月以来，1P API 的使用变得更加集中，前 10 大 O*NET 任务目前占流量的 33%，高于 28%。另一方面，Claude.ai 的任务自我们 2025 年 11 月的数据以来已经多元化。美国低使用率州在采用速度上继续加快，尽管步伐比上一份报告有所放缓。低采用率国家则略微落后。

With this report, we can begin to trace out how various economic primitives have changed. Coursework fell as a share of usage while personal conversations increased. We also note a slight decrease in the aggregate complexity of prompts in Claude.ai, with conversations in Claude.ai exhibiting less sophisticated inputs and shorter estimated completion times.通过本报告，我们可以开始追踪各种经济基元是如何变化的。课程作业作为使用份额有所下降，而个人对话则有所增加。我们还注意到 Claude.ai 中提示的总体复杂性略有下降，Claude.ai 中的对话表现出输入不够复杂、估计完成时间较短的特点。

Overall, Claude is used for high-value, complex work that is not broadly representative of the US economy. But as the user base has grown, less remunerated tasks have become a slightly larger share of traffic. The average value of tasks, measured as the estimated wage paid to workers in occupations associated with those tasks, has declined on Claude.ai since our first report, while rising among API users. On both surfaces, users bring their most complex tasks to our more powerful model class, Opus. This inflection is stronger for API customers.总体而言，Claude 被用于高价值、复杂的工作，这些工作并不完全代表美国经济。但随着用户群的增长，报酬较低的任务在流量中所占份额略有增加。自我们第一份报告以来，Claude.ai 上任务的平均价值（以与这些任务相关的职业工人的估计工资衡量）有所下降，而在 API 用户中则有所上升。在两个界面上，用户都将最复杂的任务带给我们更强大的模型系列 Opus。这种倾向在 API 客户中更为强烈。

More experienced users tend to use Claude more collaboratively, for more work-related reasons, in more complex tasks, and with more success. This pushes back against a hypothesis we made last year that automated use may be more typical of more experienced, sophisticated users; instead, we find that the most advanced users are more likely to iterate with Claude. It’s also consistent with learning-by-doing: the more time one spends using AI, the more effective one becomes at harnessing it.经验丰富的用户倾向于以更协作的方式、出于更多与工作相关的原因、在更复杂的任务中使用 Claude，并取得更大的成功。这反驳了我们去年提出的一个假设，即自动化使用可能是更资深、更复杂用户的典型特征；相反，我们发现最先进的用户更有可能与 Claude 进行迭代。这也与“干中学”一致：使用人工智能的时间越长，利用它的效率就越高。

An alternative interpretation, of course, is that these results are driven by cohort effects or survivorship bias. Early adopters could be more technical. Those who continue using Claude could be those with tasks that it is distinctly well-equipped to do. But carefully controlled regressions rule out simple versions of this confounding, like that long-tenured users bring different kinds of tasks. Over time, we will be able to more cleanly isolate cohort and survivorship bias from learning-by-doing.当然，另一种解释是这些结果是由队列效应或幸存者偏差驱动的。早期采用者可能更具技术背景。那些继续使用 Claude 的人可能是那些任务非常适合 Claude 的人。但经过仔细控制的回归排除了这种混淆的简单版本，例如长期资历用户带来不同类型的任务。随着时间的推移，我们将能够更清晰地将队列效应和幸存者偏差与“干中学”区分开来。

These observed differences in success rates could deepen inequalities in the labor market. Economists have long noted the potential for skill-biased technological change: innovations that raise wages for high-skill workers while depressing them for others. Our analysis in this report identifies a channel through which such skill-biased transformation may already be unfolding: early adopters with high-skill tasks have more successful interactions with Claude than later, less technical adopters. These early-adopting users may simultaneously be the most exposed to AI-driven disruption and most aided by AI in these initial, augmentative waves of adoption.观察到的成功率差异可能会加剧劳动力市场的不平等。经济学家长期以来一直注意到技能偏向型技术变革的潜力：即提高高技能工人工资同时压低其他人工资的创新。我们在本报告中的分析确定了一个渠道，通过该渠道，这种技能偏向型转型可能已经展开：具有高技能任务的早期采用者与 Claude 的互动比后来的、技术性较差的采用者更成功。这些早期采用用户可能同时最容易受到人工智能驱动的颠覆的影响，并在这些最初的增强型采用浪潮中从人工智能中获得最大帮助。

Appendix附录

Available here.点击此处获取。

Data availability数据可用性

Data from this report is available here.本报告的数据可在此处获取。

Authors and acknowledgements作者与致谢

First author block*: 第一作者区块*：

Maxim Massenkoff, Eva Lyubich, Peter McCroryMaxim Massenkoff, Eva Lyubich, Peter McCrory

*Lead authors of the report*报告主要作者

Second author block: 第二作者区块：

Ruth Appel, Ryan HellerRuth Appel, Ryan Heller

Acknowledgements致谢

Tim Belonax, Keir Bradwell, Andy Braden, Dexter Callender III, Miriam Chaum, Madison Clark, Evan Frondorf, Deep Ganguli, Kunal Handa, Hanah Ho, Owen Kaye-Kauderer, Jennifer Martinez, Miles McCain, Jared Mueller, Kelsey Nanan, Tyler Neylon, Dianne Penn, Sarah Pollack, Ankur Rathi, David Saunders, Michael Stern, Alex Tamkin, Kim Withee, Jack ClarkTim Belonax, Keir Bradwell, Andy Braden, Dexter Callender III, Miriam Chaum, Madison Clark, Evan Frondorf, Deep Ganguli, Kunal Handa, Hanah Ho, Owen Kaye-Kauderer, Jennifer Martinez, Miles McCain, Jared Mueller, Kelsey Nanan, Tyler Neylon, Dianne Penn, Sarah Pollack, Ankur Rathi, David Saunders, Michael Stern, Alex Tamkin, Kim Withee, Jack Clark

Citation 引用

@online{anthropic2026aeiv5,
        author = {Maxim Massenkoff and Eva Lyubich and Peter McCrory and Ruth Appel and Ryan Heller},
        title = {Anthropic Economic Index report: Learning curves},
        date = {2026-03-24},
        year = {2026},
        url = {https://www.anthropic.com/research/economic-index-march-2026-report},
}