400K Claude Code Sessions: What the Data Means for AI Routing

Anthropic published a research paper on June 16 analyzing roughly 400,000 Claude Code sessions from 235,000 people between October 2025 and April 2026. The headline finding sounds like a labor-market story — domain expertise matters more than coding background — but the underlying data is directly operational for anyone running Claude Code at scale through a gateway, billing system, or routing layer.

What happened

The paper introduces a nine-category session taxonomy (building, fixing, testing/orchestrating, operating software, planning, exploring, data analysis, prose/documents) and measures three dimensions per session: task mode, user expertise level, and session success rate.

Key findings for operators:

• Session composition: 56% of sessions involve writing or fixing code (25% write, 26% fix, 5% testing/orchestration). 17% is operating software. 13% produces analysis or prose. The "fixing" category dropped from 33% to 19% of sessions between October and April, replaced by operating software (14% → 21%) and writing/analysis.
• Expertise amplifier: Expert users trigger an average of 12 Claude actions per prompt and ~3,200 words of output. Novice users trigger 5 actions and ~600 words. This 2.4× action multiplier and 5× output multiplier hold across every work mode and task value band. Every expertise level step adds approximately 9% more actions and 13% more output in regression controlling for mode, month, and model.
• Task value growth: Average session economic value (estimated against freelance marketplace rates) rose 27% between October 2025 and April 2026, with building, operating, and fixing tasks each growing about a third.
• Success rates: Software engineers achieve 34% verified success on coding tasks; all ten largest occupation groups land within 7 points. Domain experts reach 28–33% verified success vs. 15% for novices — a full doubling. Management occupations show the highest verified rate, slightly above software engineering. Only 5–7% of expert sessions end abandoned vs. 19% for novices, a 3× gap. The success story is experts-and-intermediates vs. novices: the intermediate-to-expert gap is modest, the novice-to-intermediate jump is where most gains concentrate.
• Division of labor: People make about 70% of planning decisions; Claude makes about 80% of execution decisions. When Claude handles planning, it takes roughly 16 actions per turn. When the user retains execution control, Claude takes about 8 actions per turn.
• Management advantage: Management occupations posted the highest verified success rate across all groups, slightly above software engineering. Anthropic interprets this as management skills — directing, delegating, specifying constraints — transferring directly to directing an agent. This is a routing-relevant signal: teams with strong management culture may get disproportionately more value from higher-tier models.

Why it matters for AI engineering teams

The session taxonomy and expertise data change how to think about provisioning and cost prediction for Claude Code deployments.

Token budget planning is expertise-dependent. A routing team that models token consumption as uniform per session will systematically underprice expert-heavy workloads. A team of domain experts (lawyers building contract tools, scientists writing analysis pipelines, senior engineers orchestrating agents) will generate sessions that are 5× larger by output volume and drive proportionally higher context accumulation. If your billing system amortizes context costs across a uniform session budget, expert users are subsidized by novice users — or vice versa.

Task mode composition changes your tier selection calculus. The shift from debugging (declining) to operating software and orchestration (growing) is operationally significant. Fixing-type sessions tend to be bounded — the task ends when the test passes. Operating and orchestration sessions tend to be unbounded, running pipelines, deploying infrastructure, or coordinating sub-agents across multiple turns. These sessions are where prompt caching pays off most, where context window size matters, and where per-turn latency directly affects developer productivity. The case for routing orchestration-heavy workloads to the most capable available model tier is stronger than for a debugging queue.

The 27% task value increase suggests the routing allocation from last year is stale. If the average session is now worth 27% more (measured in freelance equivalent), your model tier policy from six months ago was calibrated to a cheaper workload mix. Teams that route to cheaper tiers as a cost control may be making that tradeoff against tasks that have quietly become more complex.

The router and operator angle

Several of the paper's findings translate directly to routing policy decisions.

Session mode as a routing signal. The paper's nine modes map reasonably well onto routing tiers. Debugging/testing sessions with fast iteration loops benefit from low-latency, cost-efficient models. Building sessions with long context accumulation (average 2,400 words of output per turn at expert level) benefit from larger context windows and higher-quality reasoning. Operating/orchestration sessions, where Claude is coordinating sub-agents or running pipelines, are the highest-leverage use case for the most capable tier — the output multiplier alone justifies the cost difference if the task value is commensurate.

Expertise inference as a dynamic routing input. The paper's expertise classifier uses prompt phrasing, error correction patterns, and domain specificity to infer user expertise from the session transcript. These signals are available at the gateway layer: a proxy that observes the first 2–3 turns of a session can estimate whether it is trending expert or novice and adjust model selection accordingly. Routing dynamically based on session complexity is a pattern that aligns with the data — expert sessions return more per token spent on a capable model.

Prompt caching ROI scales with expertise. Expert sessions accumulate more context per turn and sustain it across more turns. The ROI of prompt caching — which is proportional to how much repeated context the cache absorbs — is higher for expert-level orchestration and building sessions than for short debugging exchanges. Teams that recently fixed their prompt caching behavior (see Claude Code 2.1.181) should weight the benefit heavily against their session composition profile.

Sub-agent depth limits interact with action chains. The paper finds that Claude takes up to 100+ actions per turn in high-expertise, high-autonomy sessions. Teams that set sub-agent depth limits (a governance control introduced in recent Claude Code releases) should model the interaction: a 5-level depth cap with 12 actions per turn per expert user implies a possible fan-out of dozens of parallel actions at the leaf level. Latency and cost governance at the routing layer needs to account for this fan-out rather than treating each user session as single-threaded.

The numbers at a glance

For routing teams building cost models, tier policies, or governance around Claude Code at scale, these are the operational numbers from the paper:

Additional benchmarks:

• SWE verified success: 34% (all 10 largest occupation groups within 7 points)
• Management occupations: Highest verified success rate, slightly above SWE
• Debugging share: Fell from 33% to 19% of sessions (Oct 2025 → Apr 2026)
• Operating/orchestration share: Rose from 14% to 21%
• Average task value: Rose 27% over the study period
• Session sample: ~400,000 sessions from ~235,000 unique users

What TheRouter users should watch or try

If you route Claude Code through TheRouter or a custom AI gateway, the session mode and expertise insights are actionable for your routing configuration:

• Audit your current session token distribution to check whether expert-heavy users are driving disproportionate context costs. If so, consider separate routing policies or token budgets for sessions that show early high-expertise signals.
• Review your model tier assignments for operating/orchestration workloads separately from debugging/testing workloads. These modes now represent a meaningfully different usage profile than they did six months ago.
• Evaluate whether your team's occupation profile (management-heavy? domain-expert-heavy?) suggests a different model tier strategy than a generic developer population would.
• Use the /docs/ section for current guidance on routing configuration, fallback policy, and token accounting for multi-session Claude Code deployments.

The research paper PDF and appendix are available from the Anthropic research page and provide session-level statistics useful for benchmarking your own deployment against the population average.