Anthropic's latest research offers a more measured view of what AI can do for productivity. The company's fourth Economic Index Report looks at how Claude performs in real use, and the central finding is simple: harder tasks may offer bigger time savings, but they also fail more often.
That finding matters because productivity forecasts depend not only on what AI could theoretically speed up, but on whether the system actually completes the work successfully. Anthropic has now adjusted its earlier estimates downward after studying one million Claude.ai conversations and one million API transcripts from November 2025, just before the release of Opus 4.5.
What Anthropic measured
The report introduces five new "economic primitives," which are basic measures Anthropic generates by having Claude analyze anonymized transcripts. These measures cover task complexity, the education level needed for user inputs and Claude's outputs, whether the use case is work, study, or personal, the AI's level of autonomy, and task success.
Task complexity is framed in terms of how long a human would need to complete the same work without AI. That gives the report a way to compare short, simple requests with longer workflows that would otherwise take several hours.
The success measure is the key change. Earlier AI productivity analysis can look overly optimistic if it counts potential time savings without subtracting failed or incomplete attempts. Anthropic's new data shows why that distinction changes the forecast.
Complex work brings a clear tradeoff
Claude's API performance declines as tasks become longer. According to the report, Claude succeeds on roughly 60 percent of API requests for tasks under one hour. For tasks over five hours, that figure falls to about 45 percent.
The 50 percent threshold appears at an estimated 3.5 hours of work. In practical terms, that means the point at which API tasks become more likely than not to fail comes well before a full workday-equivalent assignment.
Claude.ai looks different. Anthropic estimates that Claude.ai would not fall below 50 percent success until around 19 hours. The researchers connect that gap to the multi-turn nature of Claude.ai conversations, where users can revise, clarify, and steer the result over time.
This distinction is important for understanding AI productivity. A delegated API task may have less room for correction. A back-and-forth Claude.ai session can keep the human involved, which appears to help longer tasks remain viable.
Why the productivity forecast fell
Anthropic previously estimated that broad AI adoption could add 1.8 percentage points to annual US labor productivity growth. Once the company adjusted for real-world success rates, that estimate fell to roughly 1.0 to 1.2 percentage points.
The estimate falls again when bottleneck effects are included. These are parts of a job that AI cannot speed up but that still determine how fast the overall work can move. With those bottlenecks included, Anthropic estimates an impact of 0.6 to 0.8 percentage points.
That is still meaningful in Anthropic's framing. The company notes that even one percentage point annually over ten years would return US productivity growth to late 1990s and early 2000s levels. Anthropic also expects future models to achieve higher success rates.
The broader lesson is that AI productivity forecasts are sensitive to execution quality. A model that can attempt a task is not the same as a model that can complete it reliably, especially as the work becomes more complex.
Jobs may be reshaped in uneven ways
The report also examines the education level associated with Claude use. Claude is used for tasks requiring an average of 14.4 years of education, equivalent to an associate's degree. That compares with 13.2 years for all tasks across the US economy.
Anthropic describes a possible net "deskilling" effect when AI takes over higher-education tasks and leaves humans with less skilled work. The report gives travel agents as one example: planning tasks could move to AI, while humans mainly handle ticketing and payment processing.
The effect is not the same across jobs. Property managers are described as an example of "upskilling," because accounting tasks could disappear while contract negotiations and stakeholder management remain.
That makes the impact of AI adoption more complicated than a simple replacement story. Some roles may lose their more demanding tasks, while others may shift toward work that requires more judgment, coordination, or negotiation.
Prompt quality, collaboration, and adoption gaps
Anthropic found a strong link between the education level required to understand a user prompt and the education level required to understand Claude's response. The correlation coefficient exceeds 0.92 both at the country level and across US states.
In plain terms, more complex and technically precise requests tend to produce more sophisticated answers. Simpler questions tend to receive simpler answers. Anthropic says Claude adjusts its response level to match the input level.
The report also shows a shift back toward collaboration on Claude.ai. Anthropic distinguishes between "augmentation," where users work with Claude through iteration, explanation, and feedback, and "automation," where users delegate a task with minimal back-and-forth.
In August 2025, automated usage had overtaken augmented usage for the first time. That has now reversed: augmented usage rose to 52 percent, while automated usage fell to 45 percent. The share of "directive" conversations dropped from 39 to 32 percent.
Anthropic suspects product updates such as file creation, persistent memory, and customizable "skills" have encouraged more collaborative workflows. The report also notes that this does not separately track agentic usage, even though agents would fall more into the automation category.
Usage remains concentrated. The ten most common tasks account for 24 percent of all Claude.ai conversations, while code and math tasks account for 34 percent on Claude.ai and 46 percent for API usage.
Adoption patterns differ by geography. Within the US, lower-usage states are catching up relatively faster, and Anthropic estimates per capita usage could equalize within two to five years. Globally, there is no convergence: the US, India, Japan, the UK, and South Korea lead in Claude.ai usage, and a one percent increase in GDP per capita correlates with 0.7 percent more Claude usage.
Anthropic is publishing the data on Hugging Face so external researchers can study AI's economic impact. The result is a more grounded picture of AI at work: useful, increasingly collaborative, but still limited by failure rates, bottlenecks, and unequal access to the skills needed to get the best results.