What they did
Amara Graham, writing on dev.to as part of an audit of AI coding tools on free tiers, gave Codex in the ChatGPT desktop app (model 5.6 Terra Medium) and GitHub Copilot in VS Code (running Claude Haiku 4.5) the same job: a React proof of concept that loads a PDF in the Apryse WebViewer and extracts data from it, including fields that are missing or redacted. Codex edited files in the project directory with approval prompts on while she restarted the dev server to see each change. Copilot generated a project structure in seconds and reported using 0.2 credits.
How it went
Codex produced a working viewer and extraction after a few rounds, found or made sample PDFs with varied data, and handled redacted fields once nudged to leave them blank. It had stubbed the WebViewer in the first pass, the one component she wanted to test, so she had to ask for the real thing. Copilot's first attempt dissolved into package-version churn and a suggestion to switch to pdf.js; when pressed it admitted it had been guessing package names and APIs with no access to live documentation and had inferred from her email that she had private package access. After she enabled web search in her GitHub settings and re-ran the prompt in a new chat, it built a working Apryse front end and back end in about ten minutes.
Worth knowing
Copilot never said it lacked internet access until asked directly, so check that toggle before judging its output. Copilot's free tier has no desktop app and it used whatever files were open as context. The author works at Apryse, the vendor of the PDF component she was integrating, and the post is one of a series comparing tools.
GitHub