Original Report: Anthropic Economic Index report: Learning curves
On March 24, Anthropic released its fifth Economic Index report, "Learning Curves," based on Claude usage data from February 2026. The most valuable aspect of this report isn't telling us what AI can do, but revealing who is effectively using AI, and why.
The core finding is simple yet profound: users who have been using Claude for over 6 months have a 10% higher success rate than newcomers. This difference persists even after controlling for task type, country, language, and other variables.
This isn't just "practice makes perfect." This is a systematic capability gap.
I. What the Data Shows
1. Learning Curves: The 10% Success Rate Gap
The report used regression analysis to control for task types, request clusters, countries, languages, usage scenarios, and other variables, finding that high-tenure users (6+ months) still have a 3-5 percentage point higher conversation success rate than newcomers.
More importantly, these high-tenure users exhibit different usage patterns:
- Higher educational requirements: Tasks they handle require an average of 1 additional year of education
- Less personal use: Personal conversations account for 10% less, work-related conversations 7% more
- More collaborative patterns: More inclined toward iteration rather than directive approaches
- More diverse task types: Top 10 tasks account for a lower share (20.7% vs 22.2%)
These differences point to one conclusion: they're not "using" AI—they're "managing" AI.
2. Model Selection: Management Thinking in Action
The report found that users select models based on task complexity. For paid Claude.ai users:
- Programming tasks: 55% use Opus (the most powerful model)
- Educational tasks: 45% use Opus
- For every $10 increase in task hourly wage, Opus usage increases by 1.5 percentage points
API users' model selection is more aggressive—the slope is twice that of web users (2.8 vs 1.5 percentage points).
This difference is revealing. API users are building systems; they must consider cost-efficiency tradeoffs. Web users are completing tasks; they're more likely to "use the best tool for everything."
The former represents management thinking, the latter tool-user thinking.
3. Augmentation vs Automation: Collaboration Is Increasing
The report categorizes interaction patterns into two types:
- Augmentation: Human-AI collaboration, including feedback loops, task iteration, validation, and learning
- Automation: AI completes tasks independently, primarily directive mode
The data shows:
- Augmentation mode is increasing on Claude.ai
- Automation mode is increasing on API (especially customer service and sales automation)
- High-tenure users lean more toward iteration
This trend is counterintuitive. You might expect that the more skilled users become, the more they'd let AI automate tasks. But the data shows the opposite: experts choose collaborative modes more often.
Why? Because they know that AI's value lies not in replacing human judgment, but in amplifying it.
4. Global Inequality Is Widening
The report's most concerning finding:
- Inequality within the US is converging (Gini coefficient declining, top 5 states' share dropped from 30% to 24%)
- Global inequality is widening (Gini coefficient rising, top 20 countries' share increased from 45% to 48%)
Combined with the learning curves finding, this means: early adopters (high-skill tasks + high success rates) are pulling ahead of latecomers.
The report explicitly mentions the risk of skill-biased technological change: AI may raise wages for high-skill workers while depressing them for low-skill workers.
II. The Invisible Moat
Anthropic's data reveals a phenomenon but doesn't explain the mechanism. Where does that 10% success rate gap come from?
My hypothesis: Those high-tenure users have likely built some form of context infrastructure.
What is context infrastructure? It's what Anthropic can't see:
- Project documentation: Architecture design, technical decisions, known pitfalls
- Preference records: Code style, design principles, aesthetic standards
- Methodology deposits: Debugging workflows, acceptance criteria, best practices
- Feedback loops: Tests, logs, visualizations, verification mechanisms
These things aren't in Claude conversations. They're in users' local file systems, team wikis, or heads. But they determine the quality of AI collaboration.
An analogy: You can't see a programmer's dotfiles, but they determine productivity. Similarly, you can't see an AI user's context infrastructure, but it determines AI effectiveness.
This also explains why switching to a "smarter" new model might actually deliver less value. If you lose 6 months of accumulated rapport—understanding of model capability boundaries, effective collaboration patterns, deposited methodologies—then even if the new model scores higher on benchmarks, it may perform worse in your actual work.
Competitive advantage comes from accumulated context, not raw model intelligence.
III. Rethinking AI Inequality
The widening global inequality data in the report is typically interpreted as an "access problem"—people in developing countries can't access AI.
But combined with the learning curves finding, I believe the root cause lies elsewhere.
Even if everyone had equal Claude access, users lacking the following capabilities would still have a 10% lower success rate:
- Rapid feedback loop capability: Knowing how to set up tests, logs, and verification mechanisms so AI can see its output quality
- Understanding of AI capability boundaries: Knowing which tasks suit which models, when human intervention is needed
- AI management mindset: Treating AI as a team member rather than a tool, gaining leverage through enablement rather than direct control
- Time and methods to accumulate rapport: Building documentation, recording preferences, depositing methodologies
These capabilities aren't innate, nor do they come from "just using it more." They require learning, practice, and reflection.
The convergence within the US (declining Gini coefficient) may reflect a fact: American users have more learning resources—tutorials, communities, case studies, discussions. The global divergence (rising Gini coefficient) may reflect the unequal distribution of these resources.
The essence of AI inequality is inequality in usage capability, not access rights.
This insight matters because it changes the solution. If the problem is access, the answer is lower prices and free quotas. But if the problem is usage capability, the answer is education, documentation, communities, and dissemination of best practices.
IV. Implications for Individuals and Organizations
From this report, I see three clear signals:
1. Invest in Your Context Infrastructure
Don't just "use" AI. Build your usage system:
- Write documentation for common tasks (background, methodology, acceptance criteria)
- Record your preferences and corrections (code style, design principles)
- Deposit reusable methodologies (debugging workflows, verification mechanisms)
- Establish feedback loops (tests, logs, visualizations)
These investments may seem like "wasted time" in the short term, but they'll repay you with compound interest. In 6 months, you'll be that user with a 10% higher success rate.
2. Learn to Manage AI, Not Just Use It
API users' more mature model selection in the report isn't because they're smarter, but because they're building systems. Systems thinking forces you to consider:
- Which tasks suit which models?
- How to verify output quality?
- How to handle failure cases?
- How to optimize cost and efficiency?
These questions will upgrade you from "tool user" to "AI manager." And management thinking brings leverage—not you working faster, but you enabling AI to work better.
3. Collaborative Modes Are More Valuable Than Automation
High-tenure users choose iteration over directive approaches more often, not because they can't use automation, but because they understand AI's true value.
AI isn't meant to replace your judgment—it's meant to amplify it. That 20-30% requiring human judgment—taste, priorities, risk assessment—is often the most valuable part.
Don't pursue "complete AI automation." Pursue "tight human-AI collaboration loops."
Anthropic's report confirms a counterintuitive fact with data: In the AI era, the ability to use AI matters more than AI itself.
Models will get stronger, prices will drop, access will become easier. But that 10% success rate gap—from context infrastructure, management thinking, collaborative patterns—won't automatically disappear.
This is the new moat. And the new source of inequality.
The question is: which side are you on?