TL;DR
AI now authors a large share of code at some companies, but available evidence points to continuing human work in defining, reviewing and testing software. Hiring data show a tighter entry-level market, while one experiment found that developers who used AI to learn a new library scored lower on a follow-up quiz than those who coded by hand.
AI is writing a growing share of software code, but evidence available in 2026 suggests developers still need to specify what should be built and check whether the result works. A November 18 report from Thorsten Meyer AI examines company disclosures, hiring figures and research experiments as students, developers and employers weigh which coding skills to prioritize.
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Anthropic says that in May 2026, Claude authored more than 80% of the code merged into the company’s codebase. Before Claude Code launched in February 2025, the company said the share was in the low single digits. Anthropic also reported that a typical engineer merged eight times as much code per day in the second quarter of 2026 as in 2024. It cautioned that lines of code measure quantity rather than quality, making the eightfold figure “almost certainly an overstatement” of the productivity gain.
The report sets that company disclosure beside a randomized trial by METR, a nonprofit AI evaluation group. In an early 2025 trial involving 16 experienced developers, participants took 19% longer to complete tasks when using AI tools, even though they estimated that AI had made them 20% faster. METR later said speed-ups were likely larger by early 2026, but described its newer data as “only very weak evidence” of the size of any gain.
Hiring and learning data add a different concern. Indeed data cited in the report 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 found employment among 22- to 25-year-olds in AI-exposed jobs was 19% below where it would have been otherwise, according to their analysis. In an Anthropic learning 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 Skills Behind AI-Written Code
The figures point to a change in where software work happens. When an AI system drafts code, people still need to define requirements, fit changes into existing systems and review the output. The report also cites a developer survey in which 66% named AI output that was “almost right, but not quite” as their biggest frustration. That kind of imperfect draft can shift time from typing toward inspection and correction.
For learners, that shift makes basic coding knowledge useful for more than producing code unaided. Reading unfamiliar code, finding bugs, writing tests and judging security are ways to check whether an AI-generated answer is suitable. The learning experiment raises a separate concern: if learners accept AI output without working through problems themselves, they may miss practice in skills such as debugging. The experiment does not establish that every AI-assisted learning method has this effect, but it shows that tool use alone did not produce the same quiz results in that study.
For employers, the narrower share of entry-level postings may make it harder for new workers to gain experience. The cited figures describe job postings and employment patterns; they do not establish that AI caused every change. Still, they put pressure on companies and training programs to explain how beginners can build practical skills when fewer listed openings are explicitly junior.
Products Worth Considering

Amazon Kindle Scribe Colorsoft 64GB with Premium Pen

Anker SOLIX C300 Portable Power Station, 288Wh

Sony WH-1000XM6 Wireless Noise-Canceling Headphones, Black
As an affiliate, we earn on qualifying purchases.
From Code Generation to Review
In March 2025, Anthropic chief executive Dario Amodei told the Council on Foreign Relations that AI could be writing 90% of code within three to six months and “essentially all of the code” within twelve months. He also said programmers would still need to specify what was being built, how it fit with other code and whether the design was secure. The later Anthropic figures offer one company-specific measure of AI-authored code, but they do not show that the broader industry reached Amodei’s forecast on that schedule.
Other figures in the report suggest adoption varies. Google chief executive Sundar Pichai said in October 2024 that AI generated more than a quarter of new code at Google, which engineers then reviewed and accepted. METR cited an estimate in February 2026 that about 4% of GitHub commits were authored by Claude Code. These measures cover different companies, dates and definitions, so they should not be treated as directly comparable industry-wide shares.
The report also notes that US software development postings were up about 23% year over year in November 2026, after reaching a low in May 2025. That rebound sits alongside the small entry-level share reported for the first quarter. The figures describe different points in time and do not establish how many jobs are available to new graduates specifically.
“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
How Far the Evidence Reaches
Anthropic’s code figures come from the company developing Claude and describe its own engineering work. They do not establish how much code AI writes across other employers or how much additional code represents useful output. The eightfold measure is based on lines merged per engineer, and Anthropic itself warned that it may overstate productivity gains.
The studies cited also differ in design and timing. METR’s randomized trial involved 16 experienced developers working on selected tasks in early 2025; its later comments on speed gains were tentative. The Anthropic library-learning experiment involved mostly junior developers, but the provided material does not specify enough detail to assess how broadly its quiz result applies. It remains unclear how these results generalize across skill levels, tools, workplaces and types of software.
The hiring data show a small entry-level share in one quarter and a year-over-year posting increase in November, but those measures do not explain why employers changed their listings. The cited Stanford finding also does not, on its own, show that AI caused the employment difference. The available information does not settle whether demand for junior developers will recover or how employers will redesign early-career roles.
What Developers Should Watch
Future evidence will need to show whether AI-assisted work improves outcomes such as reliable software, fewer defects and shorter delivery times, rather than just increasing code volume. More trials across different experience levels and real workplace tasks could clarify when AI saves time and when review or rework offsets the gain. Anthropic’s own caveat makes output quality a central measure for interpreting its reported increase in code.
Students and career switchers can watch for changes in entry-level postings and employer training practices. Developers can also assess whether AI tools help them understand unfamiliar code, test assumptions and catch errors, or whether they are accepting drafts without enough review. The current evidence supports keeping those skills in view, while leaving open how quickly job requirements and hiring patterns will change.
Key Questions
Does the report say learning to code is no longer useful?
No. It argues that skills such as reading, testing, debugging and reviewing code remain useful as AI generates more drafts. It does not claim every learner needs the same level of coding expertise.
What does Anthropic’s 80% figure measure?
Anthropic said Claude authored more than 80% of the code merged into its own codebase in May 2026. That is a company-specific measure, not a share for the software industry.
Did AI make developers faster in the METR trial?
In METR’s early 2025 randomized trial, 16 experienced developers took 19% longer with AI, despite believing they had become faster. METR later said newer data suggested gains may have grown, while describing that evidence as weak.
Are entry-level software jobs disappearing?
The cited Indeed data show that 4.5% of US software development postings were entry-level in the first quarter of 2026. That figure alone does not establish why the share was low or whether it will continue.
Source: Thorsten Meyer AI
