Back to Engineering
The Pragmatic Engineer

Cursor AI Coding: Power Users and Acceptance Rates

Cursor's AI coding tool shows power users generating 10x more code, with nearly half of AI changes accepted without review.

2 min read·Curated & commentary by AWS News Bot
aicodingdeveloper-productivitycursordeveloper-experience

Editorial summary and commentary based on the original from The Pragmatic Engineer. Read the original

Power users generate 10x as many lines of code vs the median, with almost half of AI changes accepted without manual review by devs.

What changed

  • Cursor's AI coding assistant reports that its most active users generate approximately 10x more lines of code compared to median users.
  • The majority of AI-related cost for Cursor users stems from input tokens, not output tokens.
  • Nearly 50% of AI-generated code changes are accepted by developers without subsequent manual editing.

Why it matters

This data from Cursor suggests a significant productivity delta for power users of AI coding tools, potentially indicating a new tier of developer output. The statistic on acceptance rates is particularly striking: 47% of AI changes are accepted without manual review, implying a high degree of trust or a significant reduction in the perceived cost of review for these suggestions. This could signal a shift towards developers acting more as AI orchestrators than traditional coders, particularly for well-defined tasks. The cost insight, that input tokens are the primary driver, is crucial for understanding the economic model of such tools at scale.

The catch

The honest version: This data is specific to Cursor's user base and their AI tool. It is not a universal statement about all AI coding assistants or all developers. The high acceptance rate might also reflect the specific types of tasks these power users are tackling, which could be more amenable to AI assistance, rather than a generalizable improvement across all coding domains. What this replaces: Traditional metrics for developer productivity, which may not account for the speed of AI-assisted code generation and acceptance. It also highlights a potential future where code review becomes a more targeted, rather than comprehensive, activity.

Ship it

If you are using AI coding assistants, track your own acceptance rates and compare them to Cursor's figures. Pay attention to the token costs, particularly input token usage, as this will likely become a significant factor in operational expenditure for AI-intensive workflows. Consider how your own team's

Keep reading

Related articles

Picked by tag overlap — same services and topics, different angles.

2 min read
AWS News Bot

EC2 G7e Instances Expand to EU and Asia, Boosting AI Inference

NVIDIA RTX PRO 6000-powered EC2 G7e instances are now available in Frankfurt, Stockholm, and Mumbai, offering up to 2.3x inference performance.

ec2g7envidia+3
Read
1 min read
AWS News Bot

Copilot code review: Unix tools over bespoke

GitHub's engineers found that specialized AI tools for code review increased costs. The fix? A return to simpler, shared primitives.

copilotcode-reviewdeveloper-tools+2
Read
1 min read
AWS News Bot

Anthropic's Claude Model Changes: Capacity or Strategy?

Capacity constraints or a strategic shift? Anthropic's recent Claude model changes raise questions for developers.

anthropicclaudeai+3
Read