Anthropic’s April 2025 analysis found that conversations on Claude Code were more often directive—delegating work with little interaction—than conversations on Claude.ai. But the same study also points to feedback and iteration as important parts of software work: AI doing much of a task did not mean developers’ review disappeared.
What Anthropic measured
Anthropic analyzed 500,000 interactions, split evenly between Claude Code and Claude.ai. The conversations were collected from April 6–13, 2025, and processed with a privacy-preserving tool intended to distill them into higher-level, anonymized insights. Claude.ai conversations came from Free and Pro users; sampled Claude Code sessions were powered by Anthropic’s first-party API. The analysis excluded Team and Enterprise use and other API traffic, so it does not represent all Claude use in professional settings. Anthropic’s report describes the sample and method.
As an Amazon Associate I earn from qualifying purchases.
The study classified interactions by collaboration pattern. Its categories distinguish how work unfolded, not simply whether AI contributed:
Recommended Free Tools
- Directive: the user delegates a task with minimal interaction.
- Feedback Loop: the model’s work is guided by feedback from an environment, such as a tool or execution result.
- Task Iteration: the user and model refine the result collaboratively.
- Learning: the interaction is oriented toward understanding or acquiring knowledge.
- Validation: the user seeks a check or confirmation.
That distinction matters: a task can involve substantial automation and still include human feedback, refinement, or validation.
#1 Best Overall
How Claude Code differed from Claude.ai
In the April 2025 sample, Anthropic classified 43.8% of Claude Code conversations as Directive, compared with 27.5% of Claude.ai conversations. This is a comparison of the shares of conversations in the sampled product surfaces—not a measure of the share of all coding work automated, nor a count of developers who stopped reviewing output.
The report’s software-versus-non-software comparison on Claude.ai adds a different perspective. For software development, Feedback Loops were higher by 18.3% and Directive behavior lower by 11.2% relative to non-software use. These are the report’s relative comparison figures; they should not be read as percentage-point changes. Anthropic interprets the pattern as evidence that coding involves more human review and iteration than non-coding tasks, even when Claude performs much of the work.
| Comparison | What the report found | How to read it |
|---|---|---|
| Directive conversations: Claude Code vs. Claude.ai | 43.8% vs. 27.5% | Shares within the April 2025 sample; Claude Code was more often classified as minimal-interaction delegation. |
| Software vs. non-software use on Claude.ai | Feedback Loops +18.3%; Directive behavior −11.2% | Relative differences reported by Anthropic, not percentage-point changes. |
What the findings do—and do not—say about developers
The strongest conclusion is about interaction patterns in this particular Claude sample: Claude Code conversations were more often directive, while software-development use also showed a stronger role for feedback loops than non-software use on Claude.ai. Those observations can coexist. A developer may delegate implementation and still inspect results, respond to errors, refine a solution, or validate it.
The figures do not show that AI has eliminated developers’ review work, that a particular fraction of software jobs is automated, or that all coding assistants behave this way. The study observed conversations, not the full context in which responses were used or the ultimate quality and adoption of the resulting code. Anthropic also cautions that project-type estimates depend on uncertain inferences because it does not know the real-world context of each response. The authors say software development may offer an early indicator, but its lessons should not automatically be generalized to other occupations.
- Product coverage is limited: Claude.ai data came from Free and Pro, not Team or Enterprise; Claude Code sessions were limited to those using Anthropic’s first-party API.
- Other channels are absent: the analysis excluded other API traffic, and Claude Code can also be used through third-party cloud-provider APIs.
- Classification is not a work-outcome audit: collaboration labels describe interaction patterns, not whether generated code was accepted, deployed, or correct.
- The sample is time-bound: the conversations were collected over one week in April 2025 and should not be treated as a current usage census.
How later Economic Index reports fit
Later reports provide context, but they do not directly replicate the April 2025 comparison. Anthropic’s January 2026 report, based on November 2025 interactions, introduced economic primitives and estimated software-development task success at 61% using classifier-derived measures. Its meaning is therefore different from the 2025 collaboration categories, and the authors note limits to interpreting those measures. Read the January 2026 report.
In a February 2026 sample described in a March report, Computer and Mathematical occupation tasks accounted for 35% of Claude.ai conversations; Anthropic also described coding as the most common use on its platforms and noted coding activity moving toward first-party API traffic. This is a separate sample and report, not an update to the April 2025 software-development comparison. Read the March 2026 report.
Anthropic’s June 2026 report used an autonomy framework to compare Claude Code, chat, and Cowork. It found higher measured autonomy for Claude Code on 26 of 31 output types; for scripts and code snippets, Claude Code sessions averaged 0.53 more autonomy points on a 1–5 scale. Those results concern a different construct from the 2025 categories of Directive, Feedback Loop, and Task Iteration, so the numbers should not be combined. Read the June 2026 report.
How to use the April 2025 result
Use it as evidence that, in Anthropic’s sampled interactions, Claude Code was more often used for directive delegation, while coding on Claude.ai showed relatively more feedback-loop behavior than non-software use. Do not use it as a standalone estimate of industry-wide automation, job displacement, or the amount of human oversight required in a software team. Its value is narrower: it offers a snapshot of how people collaborated with two Anthropic product surfaces during a defined week, with important gaps in coverage and context.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
