Anthropic Claude Code Usage Report June 2026 – Findings on Agentic Coding, Expertise, and Labor Implications

TL;DR

Anthropic examined roughly 400 k Claude Code sessions from October 2025 to April 2026 and discovered that users make most planning decisions while Claude handles execution, success rates match those of professional software engineers across occupations, and higher domain expertise dramatically improves outcomes and task value.

Overview of the Study

Anthropic introduced a privacy‑preserving framework to analyze interactive, agentic coding sessions. The dataset covers ~400 k sessions from ~235 k users, captured via Claude Code’s CLI, web UI, and desktop app. The analysis focuses on three dimensions: work mode, decision division, and session success, and links them to user expertise and occupation.

Work Modes and Task Composition

Conclusion: Writing‑related work dominates Claude Code usage, but non‑coding activities are growing rapidly.

  • Sessions are classified into nine mutually exclusive work modes: building, fixing, testing, orchestrating (code‑centric), operating (deployment/monitoring), understanding and planning (analysis of existing systems), and analyzing and communicating (data or prose).
  • Distribution (Figure 1): 56 % of sessions involve code creation, modification, or testing (25 % building, 26 % fixing, 5 % testing/orchestrating). Operating software accounts for 17 %, planning/exploration 14 %, and analysis/communication 13 %.
  • Classification reliability is high: >90 % of sessions labeled as code‑changing show corresponding telemetry.

Division of Labor Between User and Claude

Conclusion: Users decide what to build (≈70 % of planning decisions) while Claude decides how to build it (≈80 % of execution decisions).

  • Decision attribution: A classifier parses transcripts to label each decision as planning or execution and attributes it to the user or Claude.
  • Average shares: Users make ~70 % of planning decisions but only ~20 % of execution decisions.
  • Action count per turn: Typical sessions have ~4 prompt‑action cycles. Each user prompt triggers ~10 Claude actions (≈2 400 words of output). When Claude controls planning, actions per turn rise to ~16; when the user retains execution control, actions drop to ~8.
  • Figure 2 visualizes the planning vs. execution decision split.

Expertise Levels and Model Output

Conclusion: More expert users elicit longer, richer Claude output—up to 12 actions and 3 200 words per prompt for experts versus 5 actions and 600 words for novices.

  • Expertise is inferred from transcript cues (precision of instructions, verification requests, correction direction) on a 5‑point novice‑to‑expert scale.
  • Table 1 (Figure 3) shows exemplar novice vs. expert sessions from the public SWE‑chat dataset.
  • Action and word counts increase monotonically with expertise; regression controls confirm significance (p < 0.001) and a +9 % actions / +13 % words per expertise level.

Who Uses Claude Code?

Occupation Distribution

Conclusion: While software‑related occupations are the largest user group, a diverse set of professions—including business, arts, management, and sciences—regularly employ Claude Code.

  • Occupations are inferred from transcript context (file names, referenced artifacts, domain vocabulary) and mapped to 23 SOC categories.
  • Occupation is identifiable in ~70 % of sessions. The top categories: Computer & Mathematical (largest), Business & Financial Operations, Arts/Design/Media, Management, Life/Physical/Social Sciences.
  • Fastest‑growing non‑software groups: management, sales, legal.

Evolution of Work Types

Conclusion: Over the seven‑month window, the share of fixing sessions fell from 33 % to 19 %, while operating, writing, and analysis sessions roughly doubled.

  • Figure 4 tracks mode shares and shows a shift toward end‑to‑end agentic tasks (deployment, data analysis, document generation).
  • Estimated economic value (derived from freelance marketplace rates) rose 27 % overall, with building, operating, and fixing tasks each gaining ~30‑40 %.

Success Metrics and the Role of Expertise

Conclusion: Higher user expertise strongly correlates with verified success; the biggest gain occurs when moving from novice to intermediate.

  • Success definition: A two‑stage classifier first judges whether the user achieved their goal (success, partial, fail, no clear goal). A separate verifier looks for hard signals (git commits, pull‑request merges, passing tests, explicit user affirmation). Verified success requires both.
  • Sessions with no clear goal (~7.7 % of data) are excluded from success analysis.
  • Success rates (Figure 5):
    • Novice: 15 % verified success, 77 % at least partial success.
    • Intermediate/Expert: 28‑33 % verified success, 91‑92 % partial success.
  • In sessions that encounter trouble (failure signals > 3), verified success rises from 4 % (novice) to 15 % (expert). Partial success climbs from 60 % to ~80 %.
  • Abandonment (failed + zero lines of code) occurs in 19 % of novice sessions vs. 5‑7 % for higher expertise levels.

Occupation vs. Expertise

Conclusion: Occupation matters less than expertise; success rates across occupations differ by only a few percentage points.

  • Verified success in all sessions: software‑related users 30 %, other occupations 26 %.
  • In code‑producing sessions, verified success is 34 % (software) vs. 29 % (others); partial success exceeds 88 % for both.
  • Figure 6 shows that the ten largest occupational groups all fall within a 7‑point band of software engineers. Management occupations slightly outperform engineers on verified success, possibly due to clearer confirmation language.

Implications for the Labor Market

Conclusion: Agentic coding reduces the importance of formal coding backgrounds but amplifies the value of domain expertise; modest expertise yields most of the benefit, while deep mastery adds only marginal gains.

  • Users with strong domain knowledge can direct Claude to perform more work per instruction, leading to higher success and higher‑value tasks.
  • The division of labor suggests that coding agents act as implementation engines while humans provide problem framing and domain judgment.
  • If future models lower the expertise barrier, we may see broader adoption of software production across all occupations, reshaping skill demands.

Limitations and Future Work

  • Real‑world outcomes (e.g., post‑deployment usage of generated code) are not measured.
  • Non‑interactive Claude Code usage (e.g., batch generation) is excluded; future frameworks will capture this activity.
  • All classifications rely on model‑based transcript analysis; while telemetry cross‑checks show high agreement, large‑scale human validation remains challenging.
  • Ongoing monitoring will track whether the returns to expertise diminish as models become more autonomous.

References

All figures referenced are reproduced from the original Anthropic report.

Sources

Related

  • Dispatch
    Anthropic Measures AI Agent Autonomy in Practice – Key Findings and Implications

    Anthropic’s February 2026 study reveals that Claude Code agents are running autonomously for longer, experienced users grant more auto‑approval yet interrupt more, agents pause for clarification more than humans interrupt, and risky‑domain usage remains limited but emerging.

  • Dispatch
    Anthropic Economic Index Cadences Report – June 2026

    Anthropic’s June 2026 Economic Index report reveals that Claude usage now follows daily and weekly work cadences, produces higher‑value artifacts that consume more compute, and that users who automate more tasks are surprisingly more optimistic about AI’s impact on their jobs.

  • Dispatch
    Anthropic Economic Index: AI's Impact on Software Development

    Anthropic's analysis of 500,000 coding interactions reveals that specialist AI agents like Claude Code drive significantly higher automation rates than general chatbots, particularly in user-facing web development and among startups.

  • Dispatch
    Anthropic Economic Index March 2026 Report: Learning Curves

    Anthropic’s March 2026 Economic Index report shows that Claude usage diversified, lower‑wage tasks grew, and high‑tenure users achieve higher success rates, indicating learning‑by‑doing and potential skill‑biased impacts.

  • Dispatch
    Anthropic Economic Index January 2026 Report – Economic Primitives

    Anthropic’s January 2026 Economic Index report introduces five new “economic primitives” that quantify how Claude is used, revealing geographic concentration, task complexity, skill bias, and productivity implications.