🔍 Read the full analysis: Coding Skills In The Age Of AI: What Students, Developers And Managers Can Take Away on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Anthropic says Claude authored more than 80% of code merged into its own codebase in May 2026, but that company-specific figure does not show that AI has raised productivity by the same amount. Research finds mixed results on developer speed, fewer entry-level software job postings and a learning trade-off when students rely on AI. The evidence supports learning to specify, review, test and debug code, not abandoning programming skills.
A November 18 review of employment data, developer experiments and company disclosures finds that AI is writing a growing share of some software code, but the evidence does not show that programming skills are no longer worth learning. Anthropic says Claude authored more than 80% of code merged into its codebase in May 2026; separate studies show uncertain productivity effects and potential learning costs when people rely on AI.
Anthropic’s own figures are among the clearest measures of AI-generated code, but apply to one company that builds and uses the tool. The company says Claude authored more than 80% of its merged code in May 2026, compared with a share in the low single digits before Claude Code launched in February 2025. It also says its typical engineer merged eight times as much code per day in the second quarter of 2026 as in 2024.
Anthropic cautions that lines of code measure quantity rather than quality, making the eightfold figure “almost certainly an overstatement” of productivity gains. In a March 2026 employee poll, the median estimate was around four times as much output with AI; Anthropic said it expected the true uplift to be lower. Those are company estimates, not an independent measure of output across the software industry.
Controlled tests offer mixed results. In a 2025 randomized METR trial, 16 experienced open-source developers completed 246 tasks in familiar projects. Tasks took 19% longer when AI tools were allowed, although participants believed the tools had made them 20% faster. By contrast, a 2022 study of 95 developers found that participants using GitHub Copilot completed a small JavaScript server task 55.8% faster. The different tasks and tools limit direct comparison.
Skills Beyond Code Generation
The practical question is not just how much code a model can produce, but whether people can describe the right task, check the result and catch failures. Amodei’s account of programming puts those responsibilities with the human: the programmer must specify what is being built, how it fits with existing code and whether the design is secure.
That distinction matters for students and career changers facing a tighter entry-level market. Indeed data cited in the review show that 4.5% of US software development postings were entry-level in the first quarter of 2026, while 69.3% were senior. Stanford researchers, as summarized in the source, estimate employment among 22- to 25-year-olds in AI-exposed jobs was 19% below where it would otherwise have been. These figures indicate pressure, but do not establish that AI alone caused the change.
For employers, output measures need care: more generated code does not automatically mean better software or less work. Managers still need people able to review, test and maintain systems. For learners, the evidence suggests using AI as an aid while practicing the underlying skills needed to spot incorrect or insecure output.
Products Worth Considering

Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black

Amazon Kindle Scribe Colorsoft 64GB with Premium Pen

Samsung 32 Odyssey G5 QHD Gaming Monitor
As an affiliate, we earn on qualifying purchases.
Mixed Signals From Work and Learning
The evidence covers different settings and should not be treated as one industry-wide productivity estimate. Google chief executive Sundar Pichai said in October 2024 that more than a quarter of new code at Google was generated by AI and then reviewed and accepted by engineers. METR cited an estimate in February 2026 that Claude Code authored about 4% of GitHub commits. These numbers use different populations and measures from Anthropic’s internal codebase figures.
METR’s later attempt to measure changing tools was inconclusive. A second study begun in August 2025 produced an “unreliable signal,” the research group said, partly because some developers were reluctant to take tasks they did not want to do without AI. Raw estimates suggested tasks took 18% less time for developers from the first study and 4% less for newly recruited developers, but both confidence intervals included no effect. METR said the results were only very weak evidence about the size of any improvement by early 2026.
Learning outcomes add another qualification. In an Anthropic experiment, mostly junior developers who used AI while learning a new library scored 50% on a follow-up quiz, compared with 67% among those who coded by hand. The largest gap was in debugging. The experiment is a specific learning test, not proof that every use of AI harms learning, but it shows why producing working code and understanding it are not the same outcome.
“The programmer still needs to specify what is being built, how it fits with other code, and whether this is a secure design or an insecure design.”
— Dario Amodei, Anthropic chief executive, speaking at the Council on Foreign Relations on March 10, 2025
Productivity and Hiring Remain Unsettled
It is not yet clear how much AI changes software productivity across employers. Anthropic’s internal code counts, a small controlled trial and developer self-reports measure different things; none alone establishes an industry-wide effect. METR’s later estimates remain uncertain, and the source does not provide a common method for comparing the studies.
The job figures also do not establish that AI caused the reduction in entry-level postings or the employment estimate for younger workers. The review reports that US software development postings were about 23% higher year over year in November 2026 after reaching a low in May 2025, but the comparison does not resolve how roles are changing or whether the rebound is durable. Longer-term effects on hiring, training and software quality remain unknown.
Track Hiring and Measured Outcomes
The next useful evidence will come from broader, repeatable studies that measure task completion, software quality and maintenance costs—not just code volume—and compare similar work with and without AI. Hiring data over time will also help show whether the reported entry-level squeeze persists as software job postings change.
Until those results are clearer, students can build programming fundamentals while practicing AI-assisted review, testing and debugging. Developers can judge tools against the demands of their actual codebases, and managers can track quality and outcomes alongside generated-code counts. The current evidence supports adaptation, but not a firm prediction that AI will eliminate programming work.
Key Questions
Does the evidence show that students should stop learning to code?
No. The evidence suggests that coding fundamentals remain useful, particularly for specifying tasks, reviewing generated code, testing behavior and debugging. The learning experiment also cautions that relying on AI can leave gaps in understanding.
Does AI make developers faster?
The results are mixed. METR’s 2025 trial found experienced developers took 19% longer with AI on the tested tasks, while a 2022 Copilot experiment found faster completion of a small coding task. METR’s later estimates are uncertain and do not establish a general speed-up.
What does Anthropic’s 80% figure measure?
Anthropic says Claude authored more than 80% of code merged into Anthropic’s own codebase in May 2026. It is a company-specific figure, not an estimate of AI’s share of all software code.
Why are entry-level software jobs a concern?
Indeed data cited in the review show that 4.5% of US software development postings were entry-level in the first quarter of 2026. The figure signals a narrow share of postings, but does not show that AI caused it or predict how hiring will change.
Source: ThorstenMeyerAI.com
