Is Claude 4 coming? Anthropic releases a 38-page economic index report, stating that 43% of human jobs are being replaced by AI!

Claude 3.5 Opus is gone, and Anthropic may release Claude 4 ahead of schedule this week. Netizens have revealed that, in addition to Claude 4, an inference model will also make its debut, with scores that comprehensively surpass o3. Anthropic has been silent for too long; last year, there were rumors of internal model development being hindered. Not long ago, perhaps due to the impact of DeepSeek, OpenAI and Google have been intensively launching new models. This time, Anthropic is set to make a big move: scores surpassing "full-powered" o3, which is currently recognized as a stronger competitor than DeepSeek-R1! On the same day, they also released the "Anthropic Economic Index" report, which comprehensively analyzes millions of anonymous Claude conversations over 38 pages, revealing the current application status of AI in various professions. Anthropic believes that in the coming years, AI will have a significant impact on people's work, and the latest report aims to track the long-term effects of AI on the labor market and the economy. AI Will Automate 43% of Jobs In this report, Anthropic discusses the impact of AI on different professions and groups from the perspectives of task types, usage depth, occupational skills, and income levels. The main conclusions are as follows: The use of AI is primarily concentrated in software development and writing tasks, which together account for nearly 50% of total usage. The scope of AI usage is broader, with about 36% of professions using AI in at least a quarter of related tasks; it is most common in medium to high-income jobs. Among various tasks, 57% of usage indicates that AI enhances human capabilities, while 43% indicates the replacement of human labor, i.e., job automation. It is worth mentioning that Claude's coding and mathematics account for one-third (37.2%) of all usage. Why Release the "Economic Index"? Based on millions of anonymous conversation data on the Claude.ai platform, the first report of this index provides unprecedented analysis and insights, revealing how AI can be practically applied in various tasks of the modern economy. So far, this is the clearest picture of AI's integration into real economic activities. To promote broader research, Anthropic has decided to open-source the dataset used for this analysis, allowing researchers to further explore and expand upon it. In the face of the impending transformation of the labor market and its impact on employment and productivity, effective policy measures require multiple perspectives and a comprehensive approach. Therefore, Anthropic sincerely invites economists, policy experts, and other researchers to provide valuable insights. Data Analysis Method: Linking Conversations and Occupations The new paper is based on long-term research on the impact of technology on the labor market, from the spinning jenny during the Industrial Revolution to today's automotive manufacturing robots. It focuses on the ongoing impact of AI. The new research does not investigate how people use AI nor attempts to predict the future; instead, it directly obtains data on the actual use of AI. Analyzing Occupational Tasks The new research begins with an important insight from the economic literature: sometimes, it makes sense to focus on occupational tasks rather than the occupations themselves. Jobs often share certain tasks and skills: for example, visual pattern recognition is a task shared by designers, photographers, security screeners, and radiologists. Certain tasks are more suitable for automation or enhancement by new technologies than others. Therefore, it is expected that AI will be selectively adopted across different occupations, and analyzing tasks (rather than just overall jobs) will provide a more comprehensive picture of how AI integrates into the economy. Using Clio to Match AI Use with Tasks This research is made possible by the Clio system. While protecting user privacy, it can analyze conversations with Claude. The new research analyzed approximately one million conversations with Claude (including both Free and Pro versions) and categorized the conversations by occupational tasks. The U.S. Department of Labor maintains a database containing about 20,000 specific job tasks, known as the Occupational Information Network (O*NET). Based on the classification of tasks by the U.S. Department of Labor, Clio matches each conversation with the O*NET task that best represents the role of AI in the conversation. The Clio system transforms conversations with Claude (strictly confidential, located at the top left of the figure) into occupational tasks (top middle of the figure) and further maps them to the occupational/job categories provided by O*NET (top right) Subsequently, this data can be input into various analyses (the bottom row of the figure below). Finally, tasks are grouped into the occupations they best represent according to the O*NET classification scheme, and occupations are grouped into a small number of overall categories: education and libraries, business and finance, etc. Algorithm flow summary Specific Results Impact on Different Occupations The tasks and occupations that utilize AI the most fall under the "Computer and Mathematics" category, primarily covering software engineering roles. 37.2% of Claude queries belong to this category, involving tasks such as software modification, code debugging, and network troubleshooting. The second largest category is "Arts, Design, Sports, Entertainment, and Media" (10.3% of queries), mainly reflecting various writing and editing tasks performed using Claude. Not surprisingly, occupations involving a significant amount of physical labor, such as those in the "Agriculture, Fishing, and Forestry" category (0.1% of queries), show the least representation in the data. The new research also compares the proportions in the data with the prevalence of each occupation in the overall labor market, with detailed results shown in the figure below. Degree of AI Usage Within Occupations The analysis found that heavy users of AI in the workplace are concentrated in a very small number of occupations: only about 4% of occupations use AI in at least 75% of their tasks. Moderate usage of AI is more common: approximately 36% of occupations use AI to some extent in at least 25% of their tasks. Distribution of AI applications in the economy: based on real usage data from Claude.ai. The percentages in the data represent the proportion of specific tasks, occupations, and categories involved in conversations with Claude. Distribution of AI applications in the economy: based on real usage data from Claude.ai. The percentages in the data represent the proportion of specific tasks, occupations, and categories involved in conversations with Claude. As predicted, there is no evidence in the data that occupations are fully automated: rather, AI has been widely applied to many tasks in the economy, with a greater impact on certain groups of tasks than others. AI Usage and Salaries The O*NET database provides the median salaries for various occupations in the United States. Researchers incorporated this information into the analysis to compare the median salaries of different occupations with the level of AI usage in their tasks Interestingly, both low-paying and high-paying professions have a low AI usage rate (these jobs typically involve a lot of manual operations, such as hairdressers and obstetricians). The professions with the highest AI usage are mainly concentrated in the medium to high salary range, such as computer programmers and copywriters. The relationship between annual salary (x-axis) and the percentage of conversations involving that profession (y-axis). Some representative professions are highlighted. Automation vs. Augmentation Researchers also analyzed in more detail how tasks are executed. They focused on which tasks belong to "automation" (i.e., AI directly executes tasks, such as formatting documents) and which tasks belong to "augmentation" (i.e., AI collaborates with users to complete tasks). Overall, AI tends to favor the augmentation mode in task execution, with 57% of tasks belonging to augmentation and 43% belonging to automation. In other words, in more than half of the cases, AI does not replace humans in completing tasks but collaborates with humans in activities such as validation (e.g., checking users' work), learning (e.g., helping users acquire new knowledge and skills), and task iteration (e.g., assisting users in brainstorming or performing repetitive generative tasks). The ratio of augmentation to automation involved in Claude conversations, along with a breakdown of task subtypes within each category. The 38-page report also covers some other interesting data. Figure 4 shows an analysis of the depth of AI usage across different professions. About 36% of professions use AI in at least 25% of their tasks, while only about 4% of professions use AI in 75% or more of their tasks. This indicates that, in most professions, the integration of AI is still selective rather than comprehensive. Figure 5 shows the distribution of professional skills among Claude conversation users. Skills such as critical thinking, writing, and programming appear frequently in conversations, while manual skills like equipment maintenance and installation are relatively rare. Figure 8 presents a comparative analysis of task usage patterns between Claude Sonnet 3.5 (new version) and Claude Opus models, showcasing differences in user preferences The former shows more usage in coding and technical tasks, while the latter is more used for creative writing and educational content development. Task Hierarchy Usage Status As mentioned above, researchers created a task hierarchy system using Clio to match conversations to the most suitable O*NET tasks. At the top level (Figure 11), it can be seen that: IT, technology, and related tasks dominate (accounting for nearly 50% of conversation volume) The second level consists of creative and cultural work, related to the creation and preservation of art, culture, and religious artifacts (about 20%) Business management, finance, and customer service operations constitute the third largest category (about 5%) The remaining categories each account for no more than 15% Medical services and environmental systems have low representation, each accounting for less than 5% At the middle level (Figure 12), the data reveals more refined task patterns: Software development and website maintenance are the most common activities (about 14%) Computer systems programming and debugging follow closely (about 11%) System management, hardware/software troubleshooting, and documentation publishing processes (each accounting for 4-6%) Marketing/promotion strategies, web optimization, academic tutoring, and public relations management appear but with lower frequency (each about 2-3%) Data science and machine learning applications (about 2%) At the foundational level (Figure 13), highly specific technical operations can be seen: Software modification and bug fixing activities dominate, with tasks focused on adapting to new hardware or improving performance appearing most frequently Initial debugging programs, system management, and hardware/software troubleshooting are the next most common activities Document editing and program analysis tasks appear less frequently but still constitute an important part of the conversation Note! The new research provides a unique perspective on how AI is changing the labor market. However, like all studies, it also has significant limitations. Here are some key considerations: Unclear task usage: It is unclear whether individuals using Claude to complete tasks are doing so for work. For example, someone seeking writing or editing advice from Claude may be doing so for work or for personal writing of a novel The response usage is unclear: It is unknown how users utilize Claude's responses. For example, do they copy and paste code snippets? Do they verify the responses or accept them uncritically? Limited data sources: Only the data from the free and professional versions of Claude.ai were analyzed, excluding data from API, team, or enterprise users. Task classification errors: Due to the wide variety of tasks, Clio may misclassify some conversations. Creative uses not covered: Claude cannot generate images (except indirectly through code), so creative uses are not referenced in the data. Coding use cases may be overrepresented: Since Claude is marketed as a top model for coding, coding use cases may be overrepresented in the data. Therefore, we do not believe the use cases in the dataset represent the general use of artificial intelligence. Conclusion and Future The use of artificial intelligence is rapidly expanding, and the capabilities of models are continuously improving. The landscape of the labor market may change significantly in a short period. Therefore, Anthropic will continuously repeat the above analysis to help track potential social and economic changes and regularly publish results and related datasets. This longitudinal analysis can provide new insights into AI and the job market. For example, changes in the depth of AI usage within occupations can be monitored. If AI is only used for specific tasks and only a few occupations use AI for most tasks, the future may involve the evolution of most existing occupations rather than their disappearance. The ratio of automation to augmentation can also be monitored to understand in which areas automation is becoming more prevalent. Note that the new research is merely an analysis of the conversational data from the AI model Claude and does not provide policy recommendations. Preparing for the impact of AI on the labor market cannot be directly derived from research alone but needs to be combined with evidence, values, and broad experiences. New Intelligence, original title: "Is Claude 4 Coming? Anthropic Releases 38-Page Economic Index Report, 43% of Human Jobs Are Being Replaced by AI!" Risk Warning and Disclaimer The market has risks, and investment should be cautious. This article does not constitute personal investment advice and does not take into account the individual user's specific investment goals, financial situation, or needs. Users should consider whether any opinions, views, or conclusions in this article align with their specific circumstances. Investing based on this is at one's own risk