Anthropic Cuts AI Productivity Forecasts in Half After Analyzing Claude's Real-World Failure Rates

Why It Matters

The revised forecasts temper expectations for AI’s macroeconomic impact and highlight emerging skill gaps that could shape workforce policy.

Key Takeaways

• Complex tasks halve Claude's success rates
• US productivity boost revised to under 1 percent
• Higher education prompts yield better AI answers
• Augmented usage now exceeds automated usage
• Global AI usage remains tied to GDP per capita

Pulse Analysis

Anthropic’s latest Economic Index Report provides the first large‑scale, systematic look at Claude’s real‑world performance, drawing on a million chat logs and a million API transcripts from late 2025. By quantifying task complexity and success rates, the study shows that while complex work promises larger time savings, Claude’s success drops to about 45% for tasks longer than five hours. This reality check forced Anthropic to cut its earlier estimate of a 1.8‑point productivity lift to under one point, with bottleneck effects pushing the figure to 0.6‑0.8 points. The adjustment reshapes investors’ and policymakers’ expectations for AI‑driven growth.

The report also uncovers a pronounced deskilling trend: Claude is most often deployed on tasks that require roughly 14.4 years of education, higher than the economy‑wide average. As AI automates these higher‑skill activities, workers are left with lower‑skill duties, exemplified by travel agents shifting from itinerary design to ticket processing. A striking correlation—over 0.92—between the education level of user prompts and Claude’s response quality suggests that societies with more educated workforces can extract greater value from AI, widening the gap between high‑GDP and low‑GDP regions.

Usage patterns are evolving, with collaborative augmentation now outpacing full automation for the first time since mid‑2025. Product enhancements such as persistent memory and customizable skills encourage iterative workflows, keeping users in the loop. While the United States shows rapid regional convergence in per‑capita AI usage, global adoption remains tightly linked to income levels. Anthropic’s open data on Hugging Face invites further research, and future model upgrades could improve success rates, potentially restoring some of the lost productivity upside.