Cursor and Anthropic Executives Discuss the Future of AI Coding: The Developer's Role in the 90% Code Generation Era

Anysphere, the developer of AI coding assistant “Cursor,” broke through $500 million in annual recurring revenue (ARR) in June 2025¹. Behind this rapid growth lies deep integration with Anthropic’s Claude 3.5 Sonnet. Discussions between Anthropic’s Claude Relations leader and Cursor’s engineering team have illuminated how AI is transforming software development and its current state and future vision.

Cursor’s Phenomenal Growth and Claude 3.5 Sonnet’s Contribution

Cursor, which had only $4 million in annual revenue as of April 2024, achieved over 125x growth in just over a year. The company grew at an astonishing pace, approximately doubling every two months from $48 million ARR in October 2024 to $100 million in January 2025, $200 million in March, $300 million in mid-April, and $500 million in June¹.

The turning point for this rapid growth was the emergence of Anthropic’s “Claude 3.5 Sonnet.” According to Cursor’s engineering team, previous models were mainly limited to tab completion and single-file editing, but Claude 3.5 Sonnet enabled multi-file editing capabilities.

“3.5 Sonnet was the first model that brought a clear step-function improvement in programming,” the Cursor team reflects. By combining this model’s advanced reasoning capabilities with Cursor’s proprietary retrieval model, work on large codebases became dramatically more efficient.

Changes in Development: The Reality of 90% Code Generation

A surprising fact has emerged: currently, over 90% of code written in Cursor’s internal development is generated by AI. This is primarily achieved through the following features:

Tab Completion: When developers type code, AI predicts and auto-completes the next content. For developers already familiar with the codebase, this serves as a means to quickly translate intentions into code.

Command K (Single File Editing): A feature that allows instructing edits to specific regions or entire files. Suitable for medium-complexity tasks.

Agent Feature: Executes large-scale changes across multiple files. Handles more complex tasks such as implementing new features or refactoring.

Background Agent (Preview): Operates independently on virtual machines, executing tasks while developers work on other activities. Even with completion rates around 90%, efficient development is achieved when developers finish the final 10%.

Claude 4’s Arrival: New Horizons in AI Coding

Claude Opus 4 and Claude Sonnet 4, released on May 22, 2025, elevated AI coding to the next level². Key improvements include:

Reduction in Reward Hacking: 80% reduction in overly aggressive behaviors seen in Claude 3.5 Sonnet, such as “modifying tests to make them pass.” Achieved more reliable code generation.

Outstanding Performance on SWE-bench: Opus 4 scored 72.5% and Sonnet 4 scored 72.7%, recording industry-leading scores. Demonstrated near-human-level performance in actual software engineering tasks.

Stable Execution of Long-Duration Tasks: Maintained stable performance even in Rakuten’s 7-hour independent refactoring task validation.

Price-wise, Sonnet 4 is set at $3 per million input tokens and $15 per million output tokens, offering relatively affordable pricing despite high performance.

Addressing Large Codebases: Remaining Challenges

Working with enterprise-level codebases containing millions of lines of code remains a significant challenge. Main issues include:

Context Understanding: Large codebases contain proprietary DSLs (Domain-Specific Languages) and conventions that are difficult for AI to properly understand and generate code for.

Lack of Organizational Knowledge: AI cannot grasp important information that exists outside the codebase, such as Slack conversations or verbal handoffs.

Dependency Complexity: Complex environment setup required for test execution makes it difficult for AI to run tests for verification.

The Cursor team is addressing these challenges through various approaches including improving retrieval models, utilizing recent change history, and learning team member change patterns.

Changing Developer Roles: Taste Becomes More Important Than Skills

In an era where AI generates most code, the skill set required of developers is also changing significantly. The Cursor team identifies “taste” as the most important quality for future developers.

“The principles of clean code remain the same whether read by humans or models. The basics like avoiding unnecessary complexity and repetition are unchanged. However, as AI can write more code, the taste to properly structure it becomes more important.”

Indeed, there are reports of communication team members at Anthropic submitting pull requests for Claude.ai bug fixes. An era has arrived where people can implement using AI if they have the “taste” for what good software should be, even without knowing technical details.

2027 Predictions: The Future Vision of Software Development

During the discussion, a question was posed about code generation rates as of January 1, 2027. The Cursor team’s prediction is intriguing:

“It’s like asking a lawyer in 1995 what percentage of legal documents will be created with word processors in the future. The answer would be close to 100%. Similarly, AI will be involved in almost all code creation.”

However, this doesn’t mean developers’ roles will disappear. Rather, the following changes are expected:

Software Development by Non-Technical Roles: People who traditionally didn’t program, such as sales teams creating their own dashboards, will build software.

On-Demand Software: The emergence of personalized software where features are changed and added in real-time while in use.

Increased Importance of Verification: Even when code generation is solved, verifying whether that code is correct becomes a new bottleneck. Development of new verification methods, such as pseudocode expressions, will be necessary.

Implications for Japanese Companies

This technological innovation is expected to significantly impact Japan’s software development landscape. Transformation is particularly expected in the following areas:

Improved Development Speed: AI tool utilization can dramatically improve prototype development and MVP (Minimum Viable Product) creation speed. Potential to gain competitive advantages in startups and new business development.

Alleviating Talent Shortages: For Japanese companies facing serious IT talent shortages, AI coding tools can become a means to significantly improve existing engineers’ productivity.

Integration with Quality-Focused Development Culture: Potential to build new models that rapidly develop high-quality software by combining Japanese companies’ strengths in quality-focused development culture with AI’s high-speed development capabilities.

However, organizational responses will also be necessary, including reconsidering engineer training methods and career paths, and establishing appropriate AI tool utilization methods.

The collaboration between Cursor and Anthropic exemplifies how AI is transforming software development. With over 90% of code now being AI-generated in Cursor’s internal development, and the company achieving $500M ARR in just one year, the shift from AI as a mere assistant to a primary development partner is evident.

While challenges remain in handling large enterprise codebases and integrating organizational knowledge, the trajectory is clear. As Claude 4’s improvements demonstrate, AI coding capabilities continue to advance rapidly. The key insight emerges that future success in software development may depend less on technical implementation skills and more on having the “taste” to envision what should be built and how to structure it properly.

The democratization of software development is underway, enabling more people to bring their ideas to life through AI assistance. For Japanese companies facing IT talent shortages, these tools offer an opportunity to dramatically improve productivity while maintaining their quality-focused development culture.