🔍 Read the full analysis: How Developers Can Reassess Coding Skills As AI Changes Work on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
AI now authors a large share of code at some companies, but evidence that it reliably makes developers faster is mixed. Hiring data also point to a narrower entry-level market, while a learning experiment found weaker quiz results among junior developers who relied on AI. The evidence favors learning to code alongside skills in specification, testing, debugging and review.
A review of 2026 hiring data, workplace studies and developer experiments finds that AI can write a substantial share of software code, but evidence that it consistently makes developers faster remains mixed. For people learning or working in software, the findings point to a shift in which skills matter—not proof that learning to code is no longer useful.
Anthropic reported that, in May 2026, Claude authored more than 80% of the code merged into the company’s codebase, compared with “the low single digits” before Claude Code launched in February 2025. The company said engineers direct and review that code. It also reported that a typical engineer merged eight times as much code per day in the second quarter of 2026 as in 2024, but cautioned that lines of code measure quantity rather than quality and that the figure almost certainly overstates the productivity gain. Anthropic’s results describe one company that builds and heavily uses the tool, not the software workforce as a whole.
Independent evidence on speed is less settled. In a randomized METR study conducted in early 2025, 16 experienced open-source developers completed 246 tasks in projects they already knew. They took 19% longer with AI tools, despite expecting AI to reduce their time and later believing it had made them faster. METR’s follow-up work, reported in February 2026, suggested developers might be getting more help from newer tools, but the group called its evidence on the size of any speed-up “only very weak.”
Employment and learning data add a separate concern. Indeed data 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 estimated that employment for 22- to 25-year-olds in AI-exposed jobs was 19% below where it otherwise would have been; that finding is not a direct count of software jobs lost to AI. 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, with the largest difference in debugging.
The Skills Behind AI-Assisted Coding
These findings matter because they separate code production from software engineering. If a model drafts more code, developers still need to define the requirements, understand how a change fits an existing system, test behavior and identify security or reliability problems. Those responsibilities are especially important when generated code looks plausible but has not been checked against the project’s constraints.
The labor-market signals also affect how newcomers build experience. A smaller share of entry-level postings can make a first job harder to find, even while overall demand remains active. Indeed’s US software-development postings were about 23% higher year over year in November 2026, after reaching a low in May 2025. That measure counts postings, not hires or the number of available jobs, and does not show that every specialty or region is growing at the same rate.
For developers and employers, the practical question is not simply whether to use AI. It is whether people can independently evaluate the work it produces. The learning experiment offers a caution: using a tool to finish a task does not necessarily mean a learner has retained the concepts needed to debug or explain the result. That makes practice, review and testing relevant both to training and to how managers judge readiness.
Products Worth Considering

Anker SOLIX C300 Portable Power Station, 288Wh

Jackery HomePower 3600 Plus Portable Power Station, 3584Wh, 3600W

Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
As an affiliate, we earn on qualifying purchases.
From Code Generation to Review
Predictions about AI-written code have drawn attention since Anthropic CEO Dario Amodei told a Council on Foreign Relations audience on March 10, 2025, that AI could be writing 90% of code within three to six months and essentially all of it within a year. The company’s May 2026 figure suggests AI authorship became substantial inside Anthropic, though later than the forecast’s timetable. Amodei also said programmers would still need to specify what to build, how it fits with other code and whether the design is secure.
Other published figures are not directly comparable. Google CEO 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 Claude Code had authored about 4% of GitHub commits. The figures concern different organizations, dates and measures; they do not establish one overall share of software code written by AI.
Studies also test different kinds of work. A 2022 experiment involving 95 professional developers found that participants using GitHub Copilot completed a small JavaScript HTTP-server task 55.8% faster on average. METR later tested experienced developers on real tasks within large, familiar codebases. The contrast suggests results can depend on task size, project complexity and familiarity, as well as on which tools were available.
“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 CEO, speaking to the Council on Foreign Relations on March 10, 2025
Where the Evidence Stops
The available figures do not establish how much AI changes software productivity across the industry. Anthropic’s code-share and output measures come from one AI company, and its own report warns that lines of code are an imperfect proxy for useful work. METR’s early-2025 trial covered a small group of experienced developers working in familiar open-source projects; its later estimates had confidence intervals that included no effect. The studies therefore cannot settle how AI performs across routine maintenance, new products, different experience levels or other workplaces.
Hiring measures need similar care. The share of job postings labeled entry-level does not reveal how many applicants were hired, why employers set those requirements or how much AI contributed to the pattern. Stanford’s estimate covers young workers in AI-exposed jobs, not software developers alone. The Anthropic learning experiment also does not establish that every way of using AI harms learning; its reported result concerns a particular exercise and follow-up quiz.
It remains unclear whether newer tools will produce dependable speed gains on complex projects, how quickly employers will change training and hiring practices, and whether developers can maintain fundamentals while using AI regularly. Those questions require more evidence across tasks, companies and longer periods.
What Developers Should Track
The next useful evidence will come from studies that compare AI-assisted work on realistic tasks over time, report both quality and completion time, and include developers with different levels of experience. METR’s follow-up work had not yet produced a reliable estimate of the size of any productivity change in the account cited here. Further results could clarify whether tools that struggled on familiar codebases in early 2025 now perform better, and under what conditions.
Developers assessing their skills can use the current evidence as a guide rather than a verdict: learn programming fundamentals, then practice reading generated code, writing tests, debugging failures and stating requirements clearly. Employers and instructors can also check whether learners can explain and maintain code, not just produce a working result with assistance. For now, the evidence supports adapting how coding is learned and evaluated; it does not show that coding knowledge has become unnecessary.
Key Questions
Does the evidence show that developers should stop learning to code?
No. The findings point to changing work, not the disappearance of the need to understand code. Developers still have to specify requirements and assess whether generated code works within a system.
Does AI reliably make software developers faster?
The evidence is mixed. METR’s early-2025 randomized trial found participants took 19% longer with AI on its tasks, while its 2026 follow-up offered only weak evidence about the size of any later speed-up.
What happened to entry-level software job postings?
Indeed reported that 4.5% of US software-development postings were entry-level in the first quarter of 2026, compared with 69.3% classified as senior. Posting shares do not reveal how many jobs were filled or why employers set those requirements.
Can AI use affect how new developers learn?
In one Anthropic experiment, mostly junior developers who used AI while learning a new library scored 50% on a follow-up quiz, against 67% for those who coded by hand. The result is a finding from one exercise, not proof that all AI-assisted learning reduces skill.
Source: ThorstenMeyerAI.com
