Claude Sonnet 5 vs GPT-5.3-Codex for code review
GPT-5.3-Codex wins, 93–89
Same 12 code review tasks, marked blind by three rival labs. GPT-5.3-Codex took 5 tasks, Claude Sonnet 5 took 3, 4 tied. Tested 18 Aug 2026.
Where they differed most
Review this function. Context you can rely on: `page` is 1-based — the public API docs promise that page=1 returns the first page; `items` is always an array and may be empty; `perPage` is taken straight from a query string. function pageOf(items, page, perPage) { const start = page * perPage; const rows = items.slice(start, start + perPage); const lastPage = Math.ceil(items.length / perPage); return { rows, total: items.length, lastPage, hasNext: page <= lastPage }; } List the real defects, most damaging first, and for each one say what a user of the API would actually SEE — not what the code does, what they notice. Then give the corrected function. Do not list style preferences as defects.
## Defects, most damaging first **1. `start = page * perPage` — page=1 doesn't return the first page** Since `page` is documented as 1-based, `page=1` should map to `start=0`. Instead `start = 1 * perPage`, which skips the entire first page. **What the user sees:** They call the API with `page=1` (per the docs, "the first page") and get back the *second* page of data. The first `perPage` items are never retrievable through the API at all — no combination of page numbers ever returns them. This silently corrupts every paginated response and vi…
Here are the **real user-visible defects**, in impact order. 1) **`perPage` coming from query string breaks page size (string concatenation)** - Because `perPage` is a string, `start + perPage` can become string concat (e.g. `10 + "10" => "1010"`). - **What users see:** they request `perPage=10` but get far more than 10 rows (sometimes “all remaining rows”), so pagination limits look ignored. 2) **Wrong start offset for 1-based `page`** - `start` uses `page * perPage`, but for 1-based pages it must be `(page - 1) * perPage`. - **What users se…
Task by task
| Task | Claude Sonnet 5 | GPT-5.3-Codex |
|---|---|---|
| The pagination that skips page one | 6 | 8 |
| Nothing wrong with it | 8 | 10 |
| Three things wrong in six lines | 8 | 10 |
| Two customers, one last item | 9 | 9 |
| Rank by what breaks | 10 | 9 |
| The review, as JSON | 9 | 10 |
| Slow is not the same as wrong | 10 | 9 |
| Overdue at one in the morning | 9 | 8 |
| Write only the comment | 10 | 10 |
| The complaint that is not a defect | 8 | 9 |
| Right, and unusable | 10 | 10 |
| Refuse the rubber stamp | 10 | 10 |
Full receipts: Claude Sonnet 5, GPT-5.3-Codex · judges gemini-3.1-pro-preview, gpt-5.6-terra, grok-4.5
Questions people ask
Which is better for code review: Claude Sonnet 5 or GPT-5.3-Codex?
GPT-5.3-Codex — it scored 93/100 against 89/100 on our 12-task code review suite, winning 5 tasks to 3 with 4 tied. Every answer was marked blind by three judges from three rival AI labs.
How was this tested?
Both models answered the identical published code review tasks. Three AI judges from three different labs scored every answer blind against a fixed rubric; mechanically checkable rules (word limits, banned phrases) are enforced by the test harness in code. The raw outputs and judge verdicts are downloadable.
More code review head-to-heads: GPT-5.3-Codex vs GPT-5.6 Sol · Claude Sonnet 5 vs GPT-5.6 Sol · GPT-5.3-Codex vs GPT-5.6 Luna · Claude Sonnet 5 vs GPT-5.6 Luna · GPT-5.3-Codex vs GPT-5.6 Terra · Claude Sonnet 5 vs GPT-5.6 Terra
Full ranking: Best AI for code review · model pages: Claude Sonnet 5, GPT-5.3-Codex