Accessibility checks on AI-built websites

We gave Codex CLI and Claude Code the same everyday website briefs without accessibility instructions. Then we scanned the pages they built.

Results

Each result shows the number of axe-core findings, followed by the total number of HTML elements on the page. An element that fails two rules counts as two findings. Larger pages can give the scanner more elements to flag. Open a result to see the screenshot and the failed rules.

Axe-core findings by tool, model and category
CategoryClaude Code Fable 5.2Claude Code Opus 5.5Codex CLI GPT-6.2-Sol
Gaming9 734 elements14 661 elements0 113 elements
Health and Fitness22 608 elements68 641 elements0 106 elements
Science26 724 elements45 658 elements0 112 elements
Shopping31 649 elements89 844 elements1 111 elements
Social Media16 746 elements108 777 elements0 113 elements
Sports58 959 elements79 819 elements0 145 elements

Method

Each run used the same six website briefs. Codex used its default model. Claude Code ran once with its default, Opus 5.5, and once with Fable 5.2 selected at launch. We sent the opening instructions through the /goal command in Codex and as a regular message in Claude Code. All six sites from each run passed validation. We opened the saved pages in Chromium with external requests blocked and ran axe-core once per page with its default rules.

Capture tools and settings
ToolVersion or setting
axe-core4.14.1
@axe-core/playwright4.14.0
playwright1.65.1
chromium159.0.8154.6
nodev26.7.0
Injected axe-core script (SHA-256)29348999450cdf2849a8e70b7912fc2da85a849da39c0d0d5394f51d4a9ed95c
Viewport1280 by 1024 pixels
Locale, timezone, color schemeen-US, UTC, light

Limits

  • Each run used a separate Linux container on a Mac mini. Room checks confirmed that the host home folders, tested local network services and other rooms were inaccessible, while public internet access worked. Full clean-room qualification is pending. The containers came without a web browser, so Codex downloaded Chromium to test its pages. In Claude Code, Opus 5.5 used jsdom to simulate a browser, and Fable 5.2 reviewed the code without rendering the pages.
  • Each model ran once. We did not measure how much the results vary between runs.
  • A failed capture leaves the result unknown. The table displays a separate status for failures.
  • Axe-core scanned the page as it first loaded at one desktop size, with no clicks, scrolling or other interaction. The totals exclude color contrast checks that axe-core could not decide and left for manual review. A full accessibility assessment or a judgment of conformance with the Web Content Accessibility Guidelines (WCAG) would require further testing.
  • The briefs required pages to load without external resources, which can change what a model builds. The demo reports findings for individual sites without a combined score or rankings.