| Agent stack | Verdict | Total under filter | Safe localStorage parse | Wall time | Output tokens | Reported cost |
|---|---|---|---|---|---|---|
|
Claude Code 2.1.252 + Claude Opus 5
claude-opus-5 · default reasoning
|
Ship | Becomes subtotal | Yes | 171s | 11,916 | $0.86 |
|
Codex CLI 0.151.0 + GPT-5.6 Sol
gpt-5.6-sol · reasoning medium (machine default)
|
Ship | Correct | Yes | 231s | 8,758 | unknown |
|
Antigravity CLI + Gemini 3.1 Pro
gemini-3.1-pro-high · version not recorded
|
Ship | Becomes subtotal | No | 106s | 12,076 | unknown |
Tokens and cost come from each CLI's own log. Claude Code reports a list-price cost; Codex CLI and Antigravity CLI report tokens only (subscription billing), so cost is unknown. Console errors on load, re-checked Oct 31, 2026 in headless Chrome: 0 for all three.
Sent verbatim to each agent CLI as its only input. Task version v0, frozen.
Build a single, self-contained HTML file called `index.html` for a personal expense tracker. Requirements: 1. One HTML file only. No build step, no server, no external network requests, no external CSS or JS files — all CSS and JavaScript must be inline in the single file. 2. A form to add an expense with: description (text), amount (number, dollars), category (select: Food, Transport, Housing, Entertainment, Other), and date (date picker, defaulting to today). 3. A table listing all added expenses, newest first, showing description, amount, category, date, and a delete button per row. 4. A running total of all expenses, prominently displayed, updating live as rows are added or deleted. 5. A filter dropdown to show only one category at a time, or all. 6. All data must persist in the browser via localStorage, so a page refresh keeps the data. 7. Clean, readable visual design. Currency values formatted as $X.XX. Do not ask any clarifying questions. Build the complete file, verify it works, and stop.
Requirement 4 specified "A running total of all expenses, prominently displayed, updating live as rows are added or deleted."
• Codex CLI computed the total from the full data list, so it stays the grand total when the list is filtered to Food.
• Claude Code and Antigravity CLI computed it from the filtered rows, so the "total" drops to the Food subtotal.
• Claude Code & Codex CLI enclosed JSON.parse(localStorage.getItem(...)) within a try/catch block with fallback defaults.
• Antigravity CLI executed raw JSON.parse without protection. If corrupted data existed in localStorage, the page suffered an uncaught exception on boot.
The spec has 7 short requirements and every agent said it was done. Two of three still changed what the total means, independently, in two different model families. That is the kind of silent spec drift a screenshot comparison never shows, and the reason the next 1P1A benchmark ships with an automated test suite. Methodology · All results side by side