The Anthropic Economic Index report: Economic Primitives (AI column)
An interesting one came out in January. report From Anthropic, in which they not only talked about how Claude helps everyone work more efficiently, but also shared a new approach with economic primitives. In fact, the authors took 1 Millions of Claude.ai conversations and 1 Millions of 1P API records 13-20 November 2025 And we ran them through a set of simple economic primitives -- basic coordinates of how the model is used. This allowed them to move from one large index to several signals, which allows them to speculate about the adoption of the technology. (adoption)automation (automation)susceptibility to AI operation (job exposure) and productivity. The authors selected primitives for their economic relevance, complementarity, and Claude’s ability to classify them automatically using simple prompts.
There are five new primitives, but they are actually nine classifiers. (Put them together with the prompts as pictures) 1) Task complexityHow many hours would it take for a person without AI, how many minutes it took with AI, and whether there is multitasking? 2) Human and AI skillsA person could do it without Claude, how many years of education it takes to understand prompt, and how much it takes to understand the answer. 3) Use case: work / coursework / personal. 4) AI autonomyHow many decisions were actually delegated by the models 1-5. 5) Task success: Did Claude accomplish the task? And this goes on top of the old dimension of automation vs augmentation: for example, “translate a paragraph into French” – automation is high, and autonomy is low, because there is almost nothing to decide. Engineering detail: They tried both more sophisticated classifiers and CoT tips; eventually, CoT was left only where it really improved quality – for human time, human+AI time and AI autonomy.
This classification allows you to move away from the rough metric “AI met in the task” to a multifactorial model: how difficult the task is, whether it requires rare human capital, work or study, how much a person has already given control of the machine, and whether there is any chance that automation will work not only beautifully in the demo, but also in the sale. However, this is not an external assessment, but a judgment of Claude himself.
It’s helpful for engineers to see how these primitives separate “AI for real work” from “AI as a handy little thing.” In the user sample Claude.ai software requests are pulled to 13.8 year of required education and have a success rate 61%, and questions about personal management - pull only on 9.4 years 78% success rate. At the same time, the average Claude.ai sample task without AI would take 3.1 an hour with AI-- 15.4 minutes; 46The percentage of use is work. That is, LLMs are already sitting not only in “simple” household scenarios, but occupy a significant place in the work of white collar workers.
Interestingly, the more complex and “educated” the task (More years of education)The greater the acceleration, but the lower the reliability (success rate). In Claude.ai, tasks are about the level 12 the years of education give about 9x acceleration and 16 It's about 12x. But the success rate decreases with increasing complexity: for less complex tasks, it is about 70%, and for college level scenarios about 66%. On the API, the acceleration is even higher, but the drop in success with increasing complexity is more noticeable.
Well, the last interesting insight that I wanted to talk about is the time horizon of the problem on which it gives. 50% success. If you look at the relationship between the success rate and the time it would take a person to complete a task without AI, then
- The API line 50The percentage of success is approximately 3.5 hour
- Claude.ai, about 19 clock. The authors honestly write that this is not a pure capability metric, but a mixture of model capabilities and user behavior; it is likely that multi-turn dialogue helps break down long work into manageable steps. But for tech executives, it's a sign that orchestration, feedback loops, and man-in-the-contour (human in the loop) It can change the economy of AI no less than the new benchmark model.
#AI #Economics #Metrics #Software #Engineering