AI Coding is a 100x leverage.

Author and source: GuiXing GenAI

AI coding is transitioning from "tool-driven gains" to "systemic reconstruction." Over the past two years, developers have most noticeably experienced faster code completion and reduced effort in writing boilerplate code. But now, the changes have moved beyond isolated efficiency improvements—models are beginning to read entire codebases, understand project structures, and even participate in pull requests and review processes.

Software development is shifting from "human-led, AI-assisted" to "human sets goals, AI executes processes."

This also means that the competitive dynamics within the industry are changing.

Early AI programming products competed on model capability and interaction experience—who could generate accurate results and respond quickly. But as tasks become more complex, the true competitive advantage will gradually shift to context management, caching efficiency, and cost control.

In other words, AI programming is no longer just a “smarter code editor,” but is approaching a new infrastructure for software production.

In addition, contrary to consensus, AI does not naturally narrow the developer gap; instead, it may first amplify the advantages of experts.

Developers who understand architecture, can break down tasks, and can evaluate the quality of model outputs will turn AI into a lever. Those who merely treat AI as a Q&A tool will see limited improvement.

Cursor, a company that has experienced ups and downs in the AI coding industry, has released a developer habits report that reveals the real trends shaping AI coding. It is not just another conceptual article about AI programming—it documents this transformation using actual product data.

Below is the translated text of Cursor's "2026 Spring Cursor Developer Habits Report." (Original link: https://cursor.com/insights)

A profound transformation

Software development is undergoing a remarkable transformation. This inaugural Developer Habits Report, based on real data from Cursor, documents this shift across five key dimensions:

Developers are accelerating: code writing speed has doubled year-over-year, with larger and deeper code commits, and the proportion of AI-generated code that passes review and is retained has reached an all-time high.

Smart Economics: We compared seven model families to analyze the cost per line of code and cost per commit, revealing significant differences in cost-effectiveness across models.

The advantage of top users: AI has improved efficiency for everyone, but the top 1% of developers have seen the most significant gains.

The rise of context: the amount of information the model can read has increased sharply, and the proportion of "cached reads" is also rising, enabling AI to handle more complex tasks and generate higher-quality code.

Moving toward automation: AI programming is evolving from a tool that assists individual developers into a complete system for end-to-end automated software development.

This report clearly presents with data where AI programming stands today and where it may be headed next.

Developer Accelerator

Developers work faster and produce more code, but the changes go far beyond just an increase in "quantity." AI is also transforming the "nature" of work: each code commit is larger, AI conversations involve more turns, and AI-generated code persists longer in the codebase.

1.1 Increased code output speed

The amount of code added by developers each week continues to increase, and the rate of growth has accelerated since early 2026. While lines of code are not a perfect metric, they provide a directional reference for understanding how developers' work is evolving.

1.2 The amount of code submitted each time is increasing.

The number of new code lines added per pull request (PR) has increased by approximately 2.5 times year-over-year, and the growth rate is accelerating.

1.3 Developers handling larger units of work

"Super-large commits" (defined as commits altering at least 1,000 lines of code) are becoming increasingly common as developers begin leveraging AI to accomplish larger tasks in a single commit. Notably, there was a spike in super-large commits in January 2026, when many developers were testing the latest AI programming capabilities and models. (OpenClaw)

1.4 The number of turns in AI conversations is increasing.

Over the past two months, the average number of tools invoked per AI conversation has increased by approximately 30%. The AI programming assistant is taking on more complex tasks: reading and editing files more frequently, searching code, running command-line operations, and browsing the web.

1.5AI-generated code has a longer lifespan.

Since early 2026, the percentage of AI-suggested code that remains in the codebase 60 minutes after being accepted has increased from approximately 76% to 81%.

Smart Economics

As models become more powerful, handle more context, and take on deeper tasks, their cost represents an increasingly larger share of the product experience.

To understand the trade-off between cost and intelligence, we analyzed the model’s economics from three perspectives: request cost, efficient code output, and the relationship between cost and performance scores.

2.1 The cost of requests varies significantly across different models

The cost per request varies by nearly nine times across different model series, indicating that the same workflow can have vastly different costs depending on the underlying model used.

2.2 Code acceptance rate narrowed the model price gap

The cost per request can differ by up to 9 times between inexpensive and expensive models. However, when examining the "final working code," the maximum difference is only 7 times—because the more expensive model can generate more usable code in a single request, making it less costly than it initially appears.

2.3" Cost-Quality" frontier is shifting

The chart below shows the performance of each model on Cursor's internal evaluation (CursorBench) relative to the average task cost, illustrating each model's position in terms of cost-quality trade-offs.

Superuser gap

AI is significantly boosting productivity overall, but this increase is most concentrated among the top users. The top 1% of users are gaining far greater benefits than others, and as overall AI usage grows, this gap continues to widen in absolute terms.

3.1 Super users account for the majority of AI activity

AI usage is highly concentrated: a small group of developers accounts for the vast majority of AI code lines, AI-related spending, and token consumption. The Lorenz curve illustrates this concentration, with Gini coefficients of 0.77, 0.75, and 0.72 for the three metrics respectively (on a scale from 0 to 1, higher scores indicate greater concentration of activity among a few individuals).

3.2 The output gap is widening

The absolute code output gap between the top 90% of developers and the median developer is widening, and the lead of the P99 users is even more pronounced.

3.3 The gap at the tail is even wider

This is another perspective on how the superuser gap intensifies at the tail end.

Another perspective on the staggering gap among top users: P99 developers produce 46 times as many lines of code and make 15 times as many merged commits as the median active user. While P90 users also show a clear lead, the gap is significantly smaller.

The rise of context

As models take on more complex tasks, they read more context before generating output to understand the codebase, user intent, and surrounding workflows.

This shift is cost-effective—because input tokens are significantly cheaper than output tokens, and cached token reads are even cheaper.

Improved contextual ability helps the model generate more accurate code, consistent with the increased code retention rate observed in the "Developer Acceleration" section.

Model 4.1 reads more before writing code.

The ratio of input to output tokens is rapidly increasing, indicating that the model consumes more contextual information for each token it generates. This suggests that the model performs more "preliminary work" before generating code.

4.2 The input tokens constitute the majority of non-cached tokens.

The same trend is reflected in the token composition. Input tokens now account for over 90% of the total input and output tokens, making context the dominant component in non-cached model usage.

4.3 Input context is becoming the primary source of token costs.

Input token consumption is high, but its lower unit price partially offsets its impact on costs. Even so, input tokens now constitute the majority of the "equivalent price" token cost, rising from about half at the beginning of the year to nearly 70%.

4.4 Cache reads dominate the Token

Including cache, the boundary of "context" is broader. Cache reads account for the vast majority of token activity, indicating that AI work is increasingly reliant on reusing prior context rather than re-reading all content from scratch each time.

Moving toward automation

The original purpose of AI programming tools was to accelerate the work of individual developers. We have already seen this impact in earlier chapters of this report: faster coding, larger commits, deeper AI conversation cycles, and more AI-generated code ultimately entering the commit history.

AI software development is now entering a new era—AI is becoming infrastructure, enabling end-to-end automation across the entire software development lifecycle.

5.1 More AI edits are being automatically accepted

Since the beginning of 2026, AI-generated modifications that are automatically accepted and merged into code submissions without manual line-by-line review have increased more than fivefold. This indicates that developers are placing greater trust in AI and are willing to let it handle more tasks independently within the submission process.

5.2 Automation is spreading to various workflows

Although still in its early stages, the first automated patterns are beginning to emerge. Adoption of Cursor’s automation features is growing rapidly, with security audits emerging as a strong use case for automation. Updated data also shows initial demand for the SDK execution feature—developers are seeking to transform Cursor’s AI infrastructure into a programmable platform that can be customized on demand.

Research methodology

This report is based on aggregated statistics from Cursor’s product and engineering data, including AI usage, token consumption, accepted AI code changes, and merged commit activity. Most time series charts use 7-day, 28-day, or 30-day rolling averages to reduce short-term volatility and make trend directions clearer. All metrics are reported in aggregate form to illustrate broad patterns in how developers use AI to build software. This report excludes data from users in privacy mode, including users who have signed zero-data-retention agreements with model providers.