AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: Why Students And Developers Are Reconsidering Coding In The AI Era on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

TL;DR

AI now authors more than 80% of code merged at Anthropic, according to the company, but that figure does not establish a comparable productivity gain across the industry. Research and labor-market data point to a more complicated shift: fewer entry-level postings, mixed results on developer speed, and evidence that relying on AI while learning can weaken understanding.

AI is writing a growing share of software code, but evidence available in 2026 does not show that learning to program has become unnecessary. Anthropic says Claude authored more than 80% of code merged into its codebase in May, while studies of developer productivity, entry-level hiring and AI-assisted learning show mixed results that matter to students, working developers and employers.

Anthropic reported that Claude authored more than 80% of code merged into its codebase in May 2026, compared with what it described as a low single-digit share before Claude Code launched in February 2025. The company also said its typical engineer merged eight times as much code per day in the second quarter of 2026 as in 2024. But Anthropic cautioned that lines of code measure quantity rather than quality, making the eightfold figure “almost certainly an overstatement” of the actual productivity gain. It is also a company describing its own intensive use of a tool it builds, not an industry-wide measure.

Independent evidence on speed is less settled. In a randomized METR trial involving 16 experienced developers and 246 tasks in familiar open-source projects, developers using AI tools available from February to June 2025 took 19% longer to complete tasks. They had expected AI to make them 24% faster and later believed it had saved them 20% of their time. METR said in February 2026 that a follow-up study produced only weak evidence about how much faster developers may have become with newer tools.

Hiring data show a tough entry point but not disappearing demand. Indeed data cited in the source report put entry-level roles at 4.5% of US software-development postings in the first quarter of 2026, while senior roles made up 69.3%. Stanford researchers found employment among 22- to 25-year-olds in AI-exposed jobs was 19% below where it would otherwise have been. At the same time, US software-development postings on Indeed were about 23% higher year over year in November 2026, following a low in May 2025. These figures describe different measures and do not, on their own, establish that AI caused the changes.

At a glance
analysisWhen: Based on data and studies available thr…
The developmentA review of 2026 data and studies finds that AI is changing coding work and entry-level opportunities, while the case for learning core programming skills remains.

The Entry-Level Skills Squeeze

The data suggest a shift in what employers may expect from junior programmers, rather than a clear end to programming work. If AI handles more routine code production, employers may place greater weight on whether a candidate can define a task precisely, understand an existing codebase, test results and catch mistakes. That could make the first job harder to secure: fewer entry-level postings mean fewer obvious routes for new programmers to gain experience under supervision.

There is a related risk for people learning on their own. In an Anthropic experiment, mostly junior developers who used AI to learn a new library scored 50% on a follow-up quiz, compared with 67% among those who coded by hand. The biggest difference concerned debugging. The result does not prove that all AI use harms learning, but it suggests that getting working code quickly may not build the understanding needed to maintain or repair it.

For employers, the gap between code volume and productivity also matters. More generated code can mean more review, testing and maintenance work; Anthropic itself cautioned against treating its output measure as a direct productivity measure. For students and career switchers, the practical issue is therefore not simply whether to learn coding, but whether they can demonstrate judgment beyond producing a first draft.

Products Worth Considering

As an affiliate, we earn on qualifying purchases.

What the Evidence Measures

The available studies examine different tasks and cannot be treated as one verdict on AI-assisted programming. A 2022 experiment with 95 professional developers found that participants using GitHub Copilot completed a small JavaScript HTTP-server task 55.8% faster on average. That was a short, defined exercise, unlike METR’s later trial, where experienced developers worked inside large projects they already knew, with established rules and dependencies.

That distinction helps explain why the results differ without resolving which better represents everyday work. AI may help with a contained task while adding friction in a complex codebase, where developers must check how a change interacts with existing systems. The METR trial covered tools available in early 2025; its February 2026 update said newer results were unreliable because some developers were reluctant to take on tasks without AI. METR said the evidence suggested developers were likely more sped up by early 2026, but was “only very weak evidence” about the size of the change.

Other published company figures also have different scopes. Google CEO Sundar Pichai said in October 2024 that AI generated more than a quarter of new code at Google before engineers reviewed and accepted it. METR cited an estimate in February 2026 that Claude Code authored about 4% of GitHub commits. Those measures concern different companies, periods and definitions; they should not be read as a direct industry-wide trend line.

“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 at the Council on Foreign Relations on March 10, 2025

Productivity and Hiring Remain Unsettled

It remains unclear how much current AI tools improve software-development productivity across typical teams. Anthropic’s code-volume measure comes from its own operations, and METR’s studies cover particular tasks, developers and tool versions. The results do not establish a single average effect for the industry or show how that effect changes with experience.

The hiring figures also do not isolate AI as the cause of weaker employment outcomes for young workers or the narrow share of entry-level postings. Broader economic conditions and employer decisions may also matter. The source material does not establish whether the recent increase in job postings will produce more junior openings, or whether AI-assisted learning outcomes vary substantially with how learners use the tools.

Watch for Broader Hiring Data

The clearest next signals will be new, independently measured productivity studies that test current tools on realistic work, and labor-market data that separate entry-level software jobs from senior postings over time. METR’s follow-up work may help clarify how AI affects experienced developers, though its researchers have already warned that measuring the effect has become harder as developers adopt the tools.

Students, developers and employers will also need evidence about which ways of using AI support learning rather than replace it. For now, the findings support a cautious practical approach: learn programming fundamentals, use AI as an aid rather than an unquestioned source of code, and build the ability to explain, test, debug and review what a system produces. How much those skills change hiring outcomes remains to be seen.

Key Questions

Is learning to code still worthwhile?

The available evidence does not show that coding has become pointless. It suggests the work is changing: specifying requirements, reading code, testing, debugging and reviewing design remain relevant even when AI writes code.

Does AI make software developers faster?

Results are mixed. A METR trial in early 2025 found experienced developers took 19% longer on selected tasks with AI, while a 2022 Copilot experiment found developers finished a small coding exercise 55.8% faster. Different tasks and tool versions limit direct comparison.

Are entry-level software jobs disappearing?

The cited Indeed data show entry-level roles were 4.5% of US software-development postings in the first quarter of 2026, while overall postings were about 23% higher year over year in November. These figures do not establish that AI caused the entry-level share or show whether junior hiring will recover.

Can using AI make it harder to learn programming?

One Anthropic experiment found mostly junior developers who used AI to learn a new library scored 50% on a follow-up quiz, compared with 67% for those who coded by hand. It is a specific experiment, not proof that every form of AI-assisted learning reduces understanding.

Source: ThorstenMeyerAI.com

You May Also Like

Franchise SaaS Integration Through Pre-Built MCP Connectors

A proposed salon-focused MCP connector pilot would link booking systems to AI agents, with read-only testing before write actions are built.

AI Agents On A Budget: Evaluating DeepSeek V4.1 Flash 1017

DeepSeek says its October 17 Flash update improves agent benchmarks and cuts task costs. Results are vendor-reported and not independently verified.

Building Resilient Systems With Sam Newman

In a Pragmatic Engineer podcast episode, Sam Newman discusses distributed-system resilience, microservices, observability and AI-era software work.

A Look At AutoSynthData For Enterprise Agent Training

ServiceNow CoreAI says AutoSynthData turns enterprise agents’ failures into verified training tasks; performance results and comparisons were not provided.