AI Coding Agents Are Getting Better — Does Software Become Cheaper to Produce?

Photo: Diagrams.net (objects), Juandev (code, drawing), Wikimedia Commons, CC BY-SA 4.0

01 The agents are already writing most of the new code

Start with what is measurable. Google's leadership said in late 2025 that AI generates more than 30 percent of new code at the company; Meta reports around 20 percent; Microsoft, Anthropic and Google's own product teams now describe the share of AI-written code as a routine operating metric. Claude Code crossed a billion dollars in annualized revenue within months of launch — among the fastest product ramps ever recorded. Whatever else is true, the writing of code has become dramatically cheaper for a large class of tasks.

What the headline question asks is different: does the cost of software — the thing users and enterprises pay for — fall accordingly? That is an economics question about where costs actually live, and the evidence says most of them do not live in typing.

Analysis — not prediction. N43 and Hermes AI grounds every scenario in the documented record and verified reporting as of September 19, 2026; where evidence is incomplete we say so.

02 The productivity evidence — sharper than the marketing

The experimental record is more interesting than either boosters or skeptics claim. GitHub's Copilot RCT found developers completed a standard task about 55 percent faster with the assistant. But METR's randomized trial — experienced developers on mature repositories they knew well — found agents made them about 19 percent slower: reviewing, correcting and steering the model cost more than it saved. Both results are real. They describe different worlds: greenfield tasks with clear specs, where agents are rocket fuel; and large existing codebases, where the agent is a bright intern who has not read the codebase and cannot.

The quality data is the part vendors quote least. GitClear's longitudinal analysis of hundreds of millions of changed lines finds code churn and duplicated blocks rising sharply as AI adoption grows, with copy-paste displacing refactoring. DORA's 2024 report found AI adoption associated with higher delivery throughput — and higher instability. More code, shipped faster, with quality metrics moving in exactly the direction software engineering spent fifty years trying to reverse.

03 Why cheaper code does not mean cheaper software

Software's cost structure is not typing. For any system that matters — banking, health, infrastructure, the app on your phone — the dominant lifetime costs are review, testing, security, integration, operations and maintenance. Industry estimates put maintenance alone at well over half of lifetime cost, and defects are famously more expensive to fix downstream than to prevent upstream. Agents make generation nearly free and do nothing for — arguably worsen — every other stage. The economics joke lands hard here: the code is free; the correctness is not.

This is why the honest answer to the headline is a scenario split, not a yes. Where software is small, disposable, low-stakes and spec-complete — marketing sites, internal tools, prototypes — prices will genuinely collapse, and already have. Where software is regulated, integrated or long-lived, the cost center migrates instead of shrinking: verification engineers, agent output review, AI-code audit processes. The industry is not eliminating cost; it is moving it from creators to inspectors.

04 The labor question inside the cost question

You cannot discuss software cost without discussing developers, because developer labor was most of the cost. The 2025-2026 junior hiring collapse is documented: entry-level software postings have fallen to roughly half their pre-LLM levels, and internal referrals replaced a third of them as a channel. Companies open about it — Anthropic, Google, Salesforce among them — say plainly that they are hiring fewer junior engineers because agents do that tier of work.

The unresolved tension: those junior roles were how the industry manufactured senior engineers. If nobody learns on easy tickets, review work and small features, the supply of people qualified to verify AI code — the one job the new economics needs more of — thins out a decade from now. The policy conversation is only starting to notice that AI-coding economics may be consuming its own apprenticeship pipeline, the way several skilled trades did before it.

05 The deeper question: what happens to price itself

Follow the cost curve to its end and the market-structure question appears. If marginal cost of production approaches zero in a category where prices tracked cost, prices approach zero too — and firms whose margins were built on the old cost structure do not gracefully become services companies. Some will have the distribution and data to hold prices; most will not. Consumers may capture a one-time windfall in software prices the way they did in music and media — while the production side of the industry restructures around it.

The countervailing force is Jevons' paradox: when something gets much cheaper, people buy much more of it. The evidence for elastic demand is already visible — firms are not spending less on software; they are commissioning vastly more of it, in more places, with agents. If demand elasticity exceeds productivity gains, total software spend rises even as unit prices fall. Both stories are consistent with everything observed in 2026. Which dominates will decide whether this moment reads in retrospect as software's industrial revolution or its margin compression.

06 What to watch next

Watch the first full-lifecycle cost study of AI-written codebases — not productivity per task but total cost of ownership over years; the GitClear churn data is the leading indicator. Watch where verification labor prices: if senior review rates rise while junior rates stagnate, the cost-shift thesis is confirmed in the labor market. Watch software pricing itself — the first major SaaS vendor to cut list prices citing AI production costs will signal margin compression, while seat-based pricing that quietly rises signals Jevons. And watch the insurance and audit markets: the moment AI-written code carries an insurable defect risk with a premium, we will finally have a market price for the question the headline asks — and it will be quoted in basis points, not vibes.