Anthropic has disclosed that its AI assistant Claude wrote more than 90% of the contributions to the company’s codebase in September 2026, offering a striking example of how generative AI is changing software development. The figure appeared in the company’s IPO registration filing, published on October 26, 2026. Anthropic also reported that the average daily volume of code contributed per engineer during the July–September quarter was more than 45 times the average recorded in 2024.
The figures point to a change in the work performed by software engineers at the company behind Claude. Rather than writing every line of code themselves, engineers increasingly direct AI agents, define problems, review generated code and decide which changes should be accepted. The shift is also influencing Anthropic’s hiring strategy, particularly for junior technical roles, while raising broader questions about productivity, software quality and how new developers will gain experience in an AI-driven industry.
Key Takeaways
- More than 90% of code contributions: Anthropic says Claude wrote over 90% of the contributions to its codebase in September 2026.
- Higher output per engineer: Average daily code contributions per engineer in July–September were more than 45 times the 2024 average.
- Changing engineering roles: Engineers increasingly direct and review AI-generated code rather than writing it line by line.
- Fewer junior technical hires proportionally: Anthropic’s separate employment research says junior technical hires fell from 27% to 9% of total hires since early 2025.
- AI helps build future AI: Claude is being used for research, experiments, training-data generation and model evaluations.
- Important qualification: These are company-reported figures. More code does not automatically mean better software or an equivalent increase in overall productivity.
What Did Anthropic Reveal About Claude Writing Code?
Anthropic disclosed the figures in its Form S-1 registration statement, filed as it prepares for an initial public offering.
According to the filing, Claude wrote more than 90% of the contributions to Anthropic’s codebase during September 2026. The company also reported that the average daily volume of code contributed per engineer in the three months ending September 30 was more than 45 times the average daily volume in 2024.
| Metric | Anthropic’s disclosure |
|---|---|
| AI-written code contributions | More than 90% in September 2026 |
| Code contributions per engineer | More than 45 times the 2024 daily average during July–September 2026 |
| Reporting source | Anthropic’s IPO registration filing |
| Engineering workflow | Engineers increasingly direct and review Claude’s output |
Source: Anthropic’s IPO filing coverage.
The measurements describe code contributions, not necessarily the percentage of a finished software product that Claude can independently build, maintain and ship. The filing also does not mean that human engineers have stopped participating in development.
Instead, the figures show how deeply Anthropic has integrated its own AI tools into internal software engineering.
How the Role of Software Engineers Is Changing
The traditional software development process requires engineers to write code, debug errors, run tests and revise their implementations. AI coding agents can now perform many of these steps, sometimes across complex, multi-stage tasks.
Anthropic says Claude Code has progressed from suggesting code for a human to copy and paste toward writing and running code directly. Engineers can give an agent a broader objective, allow it to work through multiple steps and then evaluate the result.
This changes where human effort is concentrated.
| Traditional engineering task | Emerging AI-assisted workflow |
|---|---|
| Writing code line by line | Describing the desired outcome and delegating implementation |
| Manually investigating bugs | Asking an agent to investigate and propose fixes |
| Running repetitive tests | Automating test execution and evaluating results |
| Reviewing every change manually | Using AI assistance for initial review, followed by human oversight |
| Choosing implementation details | Focusing more on architecture, priorities and trade-offs |
Human judgment remains important. Engineers must determine whether the proposed solution meets requirements, introduces security vulnerabilities or creates maintenance problems. They also need to understand the system well enough to identify mistakes that an AI agent or automated reviewer may miss.
The emerging role is therefore less about eliminating engineering expertise and more about using it differently.
Why Anthropic Is Producing More Code Per Engineer
Several factors may explain the increase in code contributions.
First, Claude can generate implementation code quickly. Engineers can spend less time typing repetitive structures and more time describing requirements, evaluating solutions and coordinating multiple tasks.
Second, coding agents can work through sequences of actions. They can inspect a codebase, make changes, execute tests and revise their work without requiring a person to intervene at every step.
Third, AI tools make it easier to tackle work that teams previously postponed because it was too time-consuming. This can include fixing long-standing bugs, improving test coverage and cleaning up older code.
Anthropic’s own research describes a rise in code output as Claude became capable of running code and handling longer tasks. In a May 2026 publication, the company said more than 80% of the code it merged into its codebase was authored by Claude at that point. The September figure in the IPO filing indicates that the company’s reported reliance on AI continued to increase.
However, output volume is an imperfect measure of productivity. A larger amount of generated code may require additional review, testing and maintenance. The ultimate measure of success is whether teams deliver reliable, secure software that solves the intended problem.
Anthropic Is Also Using Claude to Improve Future AI Models
The company’s reliance on Claude extends beyond ordinary application development. Anthropic says its models are increasingly involved in researching, designing and developing subsequent models—a process it describes as closing the loop.
Claude can help run experiments, generate training data, automate evaluations and support safety audits. It has also contributed to algorithmic improvements intended to make better use of computing resources.
This creates a potentially important feedback cycle: stronger models help engineers build and evaluate subsequent models, while those improvements can make future AI-assisted research and engineering more effective.
The approach could accelerate development, but it also introduces risks. Errors or weaknesses in AI-generated code and experimental processes may be difficult to identify if teams rely too heavily on automated output. Anthropic has acknowledged that AI development could progress faster than the corresponding safety and security measures.
Independent review and careful evaluation therefore remain important even when the same AI system is used throughout the development process.
What the Shift Means for Junior Software Engineers
One of the most significant implications concerns entry-level jobs.
In separate research published in October 2026, Anthropic reported that junior-level employees accounted for a smaller share of its technical workforce than in early 2025. The share of junior technical hires fell from 27% to 9% over that period.
This does not mean Anthropic stopped hiring engineers or that the same pattern applies to every technology company. Anthropic’s overall workforce grew substantially as its business expanded. The finding instead suggests that its hiring mix is changing as AI tools take on more implementation and debugging work.
For beginners, the challenge is that routine tasks have traditionally provided opportunities to learn programming, understand existing systems and develop debugging skills. If companies delegate more of these tasks to AI, entry-level employees may need to demonstrate stronger problem-solving, code-review and systems-understanding abilities earlier in their careers.
Aspiring developers should therefore learn programming fundamentals alongside AI coding tools. Understanding variables, data structures, functions, databases, APIs, testing and security makes it easier to evaluate AI-generated code rather than simply accepting it.
Source: Anthropic’s research on employment and AI.
Does Writing More Code Mean AI Has Replaced Engineers?
Not necessarily. The 90% figure describes the origin of code contributions, not the proportion of engineering jobs eliminated.
Anthropic’s own disclosures indicate that its workforce expanded while its use of Claude increased. Engineers continue to set objectives, review work and make decisions about what the company should build.
The more immediate change is in the distribution of tasks. AI can handle more implementation work, while humans increasingly focus on architecture, prioritization, evaluation and oversight.
There are also limits to the reported metrics. Code volume does not capture every aspect of software quality, and company-specific results cannot automatically be generalized to the wider industry. Anthropic has a particular incentive to use its own products extensively, and its internal workflows may differ from those of smaller businesses or companies operating legacy systems.
The more useful question is whether AI allows teams to produce better software with fewer delays and acceptable costs—not simply whether it generates more lines of code.
The Bigger Picture
Anthropic’s disclosure provides a real-world example of how AI is changing the organization of software engineering. Coding agents are moving from supplementary tools to systems that carry out substantial portions of implementation, testing and debugging. As a result, engineering teams are placing greater emphasis on defining tasks, reviewing outputs and deciding which problems deserve attention.
The transition will not affect every developer or company in the same way. Organizations must still maintain complex systems, protect sensitive data and take responsibility for software failures. But the skills valued in engineering are likely to evolve as AI takes on more routine work. The ability to understand systems, evaluate generated code and make sound technical decisions may become as important as the ability to write code quickly.
FAQs
1. Did Claude really write more than 90% of Anthropic’s code?
Anthropic’s IPO filing says Claude wrote more than 90% of the contributions to its codebase in September 2026. This is a company-reported figure describing code contributions, not proof that Claude independently developed 90% of every finished product.
2. How much did code output per engineer increase?
Anthropic reported that average daily code contributions per engineer during July–September 2026 were more than 45 times the average in 2024. The metric measures code volume and should not be interpreted as a 45-fold increase in overall productivity.
3. Is Anthropic replacing software engineers with AI?
The reported figures do not establish that Anthropic has replaced most of its engineers. The company has described a shift toward engineers directing and reviewing AI-generated work, while its workforce has also expanded.
4. Why is AI changing entry-level software jobs?
AI coding agents can handle some routine implementation, debugging and testing tasks traditionally assigned to junior developers. Anthropic’s research shows a decline in the share of junior technical hires at the company, although this does not establish an identical trend across the entire industry.
5. Should beginners still learn coding in 2026?
Yes. AI tools can generate code, but developers need enough foundational knowledge to understand requirements, identify errors, test solutions and evaluate security. Learning to use AI alongside programming fundamentals is more useful than relying on generated code without understanding it.
Looking Ahead
Anthropic’s next challenge will be to translate higher AI-generated code volume into sustained improvements in software quality, development speed and business performance. The company’s results will also help investors evaluate whether heavy internal AI use creates a lasting competitive advantage as it prepares for public-market scrutiny.
For software engineers, the key lesson is to adapt without abandoning fundamentals. AI tools can accelerate implementation, but the ability to reason about problems, review code, understand systems and take responsibility for technical decisions remains essential. The future of software development may involve less manual typing, but it will still require people who know what should be built and how to determine whether it works.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
