Three AI agents raced through a 50-page tax return
A roughly 50-page tax return, run through three coding agents side by side. Codex was accurate first time in 12 minutes, Antigravity missed a value until prompted, and Claude Code stumbled overnight.
Spotted: found in the wild. Details are as reported by the source above and have not been independently verified.
What they did
The writer’s tax return ran to roughly 50 pages, far more than a standard 1040. They had ChatGPT (Codex), Gemini (Antigravity) and Claude (Claude Code) each work through it.
How it went
ChatGPT was accurate on the first run and took 12 minutes. Gemini missed 1 value because it hadn’t loaded the full 1099 pdf, but corrected it when prompted, in 2 minutes. Claude made all manner of mistakes after an overnight run that hit its limit, though it was best at filling out the PDF forms, which ended up not mattering. They finally used Gemini in Chrome to fill the forms, and say it would frankly have been faster to copy the numbers from the spreadsheets by hand.
Try it yourself with Codex
I need to review a complex, multi-page tax return. Go through each schedule and form, flag any values that look off, and tell me how confident you are before I file, including anything you had to guess at or couldn't fully verify.
A starting prompt written by Agents in the Wild. Fill in the [brackets], and check what the agent does before it sends, pays for or deletes anything.
Possible business expenses were buried in a stack of credit card statements. Claude Code went through every statement, flagged the charges that might need reimbursing and put them in an Excel sheet.
About 15k pages of discovery stood between a solo practice and a summary judgment motion. Claude ran OCR, built embeddings and search scripts, then found evidence for each assertion, with traces to check.
Markdown files should just open, so they asked Claude Code for a viewer. It opens instantly: faster, they say, than any other macOS app, and faster than Bitwarden's Chrome extension shows its fill menu.
Migrating a password reset flow, Claude Code and Opus produced code on par with a mid level developer. It needed close supervision throughout, and cost roughly 2 months of Pro subscription for about half a day's work.
A side project needed a launch video, but not an agency's bill. Claude Code built one from the codebase plus their feedback. It took a lot of iteration over 2-3 days, all on subscription token usage.
A URL shortener side project ran on AWS. Claude Code wrote the GitHub Actions workflows for the move and fixed each failure as it came up. The writer reports zero downtime.