Five creation features shipped between April 2 and April 16, 2026. Zero maintenance features. That ratio tells you everything about where the AI coding market actually is — and what it refuses to price honestly.

Cursor 3 (April 2) launched parallel Agent Tabs — multiple AI agents writing code simultaneously in separate branches. GitHub Copilot (April 4) added Autopilot mode — agents that approve their own tool calls without asking you. Windsurf 2.0 (April 10) shipped an Agent Command Center. Claude Code (April 14) introduced Routines — scheduled background agents that generate code on triggers. OpenAI's Codex (April 16) got multi-agent workflows. Not one announcement addresses what happens to the code after it ships.

This isn't an oversight. It's a product strategy.

The 60% nobody's selling

Barry Boehm established in 1976 that maintenance consumes 60–80% of total software lifecycle cost. IBM confirmed it repeatedly in the decades since. Nothing in fifty years of software engineering has overturned that ratio.

AI appears to be widening it.

A study published on arXiv on April 8, 2026 analyzed ten large Cursor-generated projects averaging 17,000 lines of code each. Functional correctness: 91% — the code worked. CodeScene, a code health analysis tool, found 1,305 design issues across those projects: 28.4% duplicated code, methods averaging 171 lines (good practice caps at 20–30), and cyclomatic complexity — a count of branching paths through a function — averaging 17, nearly double the recommended ceiling.

The demo ships. The codebase is structurally hostile to anyone who touches it next — including the AI that wrote it.

Every major AI coding benchmark reinforces this blind spot. SWE-bench tests bug fixes. HumanEval tests function generation. No benchmark asks "can this model safely add a feature to the tangled codebase it generated three months ago with no design documentation?" Without that benchmark, vendors have zero market incentive to optimize for the thing that actually costs money.

Why AI structurally can't maintain what it creates

Maintenance demands three capabilities that creation doesn't: consistent design decisions across sessions, understanding of why code exists (not just what it does), and the discipline to refactor rather than duplicate.

Current AI tools fail at all three.

Each new session starts with zero memory of prior design choices. The model generates whatever pattern fits the current prompt, not whatever pattern it used last Tuesday. This is why the arXiv study found 28.4% duplication — the AI solves the same problem differently each time because it doesn't remember solving it before.

The METR randomized controlled trial — published July 2025, still the only controlled study of its kind — quantified the gap between perception and reality: 16 experienced developers on their own repositories worked 19% slower with AI tools, yet believed they were 20% faster. A 39-percentage-point delta between what developers think is happening and what's actually happening. On familiar codebases. On unfamiliar AI-generated code, nobody's measured it, because nobody's built the benchmark.

What a maintenance-first tool would actually require

If you designed an AI coding tool for the 60–80% of work that actually costs money, it wouldn't resemble anything shipping today.

Design rationale as metadata. Not just what the code does — why it's structured that way. Every AI-generated function should carry a record of the constraints, alternatives considered, and design decisions that produced it. Claude Code's CLAUDE.md is the closest approximation: persistent project context across sessions. But it's a text file you maintain by hand, not an automated architectural record.

Cross-session consistency enforcement. A maintenance-first tool would detect when the model introduces a pattern that contradicts an existing one and block the conflict before code generates — not after a human reviews it. Cursor's codebase indexing (up to 500MB, sub-second queries) provides the retrieval layer this requires. Retrieval without enforcement is a library without a librarian.

Refactoring as the default mode. Current tools optimize for net-new code. A maintenance tool would default to modifying existing code — locating the right place to add logic rather than generating a new file. It would measure and minimize duplication as a primary metric alongside functional correctness.

Degradation gates. When cyclomatic complexity crosses thresholds, when methods balloon past 30 lines, when duplication rates climb — the tool refuses to commit. Not as an optional plugin. As the default. The way a type checker blocks invalid code, a maintenance-first tool blocks unmaintainable code.

JetBrains' longitudinal study published April 14, 2026 tracked 800 developers across 151.9 million logged events and surfaced a signal that no coding tool currently acts on: developers spend over a third of their time verifying AI suggestions, and they reverse roughly one in five accepted completions by deleting the code later. That deletion pattern is a free training signal — a corpus of "the model thought this was right, the human proved it wrong" — sitting in every IDE's telemetry. Someone could build a feedback loop that ingests those reversals and learns to generate maintainable code from the start. Nobody has, because creation demos sell. A 30-second screencast of an agent spinning up an app from a prompt gets millions of views. A screencast of an agent carefully decomposing a 171-line method into six clean functions gets a conference talk, maybe.

The pricing reality

If you greenlit an AI-built project this quarter, here's what no vendor will quote you: budget three times the creation cost for its first year of maintenance. If your AI agent wrote 17,000 lines in a week, your engineers will spend three weeks untangling design issues before they can safely extend it — then repeat the cycle after every feature addition.

The more honest approach: treat AI-generated code as a disposable prototype with a six-month shelf life. Demo it, validate the idea, then rebuild the parts that proved their value with engineers who understand the architecture.

Every AI coding vendor in April 2026 prices and markets creation as the product. The actual cost center in software has never been creation. Until a vendor builds — and charges for — a maintenance multiplier, you're buying the cheapest phase of the process and paying full price for everything after.