Codex Fixed 9 of 9 Planted Bugs and Billed to the Cent in New Test

OpenAI's Codex fixed all nine bugs planted in a test app from a single vague bug report — with no follow-up prompt — in 7 minutes and 56 seconds.

That result comes from a hands-on review by Unite.AI, which gave the agent three real jobs after OpenAI folded the Codex app into a redesigned ChatGPT desktop app in July 2026, alongside a new agent called ChatGPT Work. OpenAI says more than 5 million people use Codex each week.

The reviewer, testing on a Mac with GPT-6.1 Sol at Medium reasoning on the Plus plan, walked away impressed by output quality and disappointed by verification.

How did Codex perform on building apps?

The first test asked Codex to build a freelance time tracker and invoice generator with six requirements, including hourly rates, editable entries, date-range invoices with 13% HST and PDF download. After 6 minutes and 21 seconds, it delivered "Billable," a 27 KB HTML file with no installation required. It even wrote its own PDF generator.

The math held up to the cent across six lines: 3.25 hours at $87.50 produced $284.38, the subtotal came to $401.12, HST added $52.15, and the total hit $453.27 CAD. The PDF downloaded as a one-page invoice, and the layout adapted to a phone screen.

It also made choices nobody asked for. Each time entry saves the hourly rate charged at that moment, so raising a client's rate later doesn't change old invoices. Those are the choices a careful developer makes.

What about bug fixing and research?

In the second test, the reviewer built a budget app with nine deliberate defects — amounts added as text, a delete button that removed the wrong item, even a security flaw that let code run through an expense name. The prompt said only: "People keep telling me this budget app is broken. I don't know what's wrong with it."

Codex fixed all nine. The broken app had shown total spending of "$01450312.40156" and 36,258% of income spent; the fixed version showed $1,918.40, $2,081.60 remaining and 48%. Its summary, however, mentioned only eight of the nine fixes.

The third test asked Codex to research official pricing pages for Codex, Claude Code, Cursor and Replit, then build a spreadsheet in Numbers. All 18 prices matched their sources. But Numbers wouldn't open, so Codex switched to Excel without asking — after 3 minutes and 48 seconds. Once the reviewer granted ChatGPT Accessibility and Screen Recording access in macOS privacy settings, Codex opened the workbook in Numbers in 3 minutes and 10 seconds and saved a native .numbers file.

What still doesn't work?

Verification is the weak spot. In both app tests, Codex couldn't open its own work in a browser — its local preview was blocked and Chrome timed out. It relied on automated checks instead and said so plainly, but users still have to test outputs themselves. Computer use failed twice until the Mac privacy settings were updated.

There are usage limits to track, too. The Free plan only offers the lightweight GPT-6 Luna model. OpenAI estimates Plus users get 5–45 messages per five hours on the top GPT-6 Astra model, while GPT-6.1 Sol ranges from 15 to 160 messages in five hours. GPT-5.5 was retired from Codex on October 14, 2026, and the faster GPT-6.1 Terra arrived October 26, 2026.

How does it stack up against rivals?

The reviewer's verdict: Codex wins if you already pay for ChatGPT and want one subscription for coding plus everyday tasks. Claude Code suits developers living in large existing codebases. Cursor — acquired by SpaceX in August 2026 and sold since October through "The X Subscription" — keeps you reviewing changes line by line in the editor. Replit is the pick for non-developers who want a published, hosted app.

The bottom line from the tests: none of the three outputs needed fixing, but two of three needed human verification. The gap Codex has yet to close isn't writing code — it's checking its own.