The Vibe Father
Roundup
WordPress is its own world. It's PHP, but it's PHP with two decades of conventions layered on top — the hook system, the loop, escaping and sanitization rules, the block editor's React-and-JSON stack, and a plugin ecosystem where "the WordPress way" often overrides whatever a language purist would do. So when someone asks which AI model is best for WordPress and plugin work, the honest first answer is: there is no WordPress benchmark. Nobody publishes a leaderboard for "wrote a correct hook and escaped the output." What we have are general coding scores plus a lot of hands-on time watching models handle real WordPress idioms. We keep the live general board at /benchmarks; here's how to read it for WordPress specifically.
The general board, and why it still guides you
The scores below are cross-language coding ability, not WordPress-specific. But WordPress work is mostly repo-surgery — editing real files inside a large, opinionated codebase — and that's exactly what SWE-bench Verified measures. So the top of this board is a reasonable proxy for "handles a plugin codebase without breaking three other things." Treat it as a starting shortlist, not gospel.
| Model | SWE | TB | LCB |
|---|---|---|---|
 Claude Fable 5 | 95.0 | 83.1 | 89.8 |
| Claude Opus 4.8 | 88.6 | 78.9 | 87.8 |
| Claude Sonnet 5 | 85.2 | – | 82.4 |
 GPT-5.5 | 80.6 | 83.4 | 85.3 |
 Gemini 3.5 Flash | 79.3 | 76.2 | 87.6 |
| GPT-5.3 Codex | 74.8 | 78.4 | 87.3 |
 DeepSeek V4 Pro | 77.6 | – | 87.5 |
A dash means the lab hasn't published that suite yet — not a zero. SWE-bench Verified is repo-surgery, Terminal-Bench (TB) is shell and tooling fluency, LiveCodeBench (LCB) is self-contained problem solving. WordPress leans on the first two.
The pick for WordPress: Claude Opus 4.8
For plugin and theme work, Opus 4.8 is our default. At 88.6 on SWE-bench Verified it handles the thing WordPress punishes weaker models for: knowing that you register a hook rather than call a function directly, that $wpdb wants prepared statements, that block registration lives in block.json, and that the loop is not just a for-loop. It also holds enough context to work across a plugin's PHP, its build config, and its block JavaScript without losing the thread. Fable 5 (95.0) is the step-up for a genuinely hard migration — say, modernizing a crusty plugin to the block editor — but for everyday plugin work Opus is the sweet spot of capability and price.
The budget pick: DeepSeek V4 Pro
A lot of WordPress work is mechanical: adding a settings field, wiring a shortcode, writing a REST endpoint, updating a template part. That volume doesn't need a frontier model. DeepSeek V4 Pro posts a capable 77.6 on SWE-bench at $0.435 / $0.87 per million tokens — roughly a tenth of the frontier rate — and it knows PHP and the common WordPress patterns well enough to grind the routine 80% for pennies. Pair it with Opus for the hard passes and your bill drops sharply. Gemini 3.5 Flash is the other value option, and its speed makes the edit-preview-refresh loop that WordPress dev lives in genuinely pleasant. The full cheap tier is in the cheapest coding models roundup.
Flag: WordPress plugins are a security surface
This matters more here than almost anywhere. WordPress powers a huge slice of the web, and plugins are the number-one source of site compromises. AI models will happily generate code that works but is insecure — unescaped output, missing nonce checks, capability checks skipped, SQL built by string concatenation. None of that shows up in a benchmark score. So no matter which model you run, every generated plugin file needs a human security pass: escape on output (esc_html, esc_attr, esc_url), sanitize on input, verify nonces on form handlers, check current_user_can() before privileged actions, and always use $wpdb->prepare(). A stronger model makes fewer of these mistakes, but none of them make zero. Our AI coding security checklist is worth pinning next to your editor.
Test it on your own stack
Because there's no WordPress benchmark, the only ranking that matters is the one you run yourself. Take a real task from your backlog — a block you need to build, a bug in a hook, a settings page — and hand the identical prompt to two or three models. Judge them on three things: did it follow WordPress conventions instead of generic PHP, did it escape and sanitize without being told, and how much did you have to correct it. That last one is the real metric. A model that's 90% right on paper but wrong about the WordPress way costs you more than a slightly lower-scoring one that gets the idioms right the first time.
What we'd run for WordPress
One model, no fuss: Opus 4.8. A two-tier setup for volume: Opus (or Fable 5 for the hard migration) driving the tricky work, DeepSeek V4 Pro or Gemini 3.5 Flash grinding the routine plugin plumbing. Whatever you pick, keep the human security pass non-negotiable and bring your own API keys so you're paying the lab's real per-token price instead of a reseller's markup — we make that case in bring your own keys. The board shifts monthly; the live version lives at /benchmarks, and if you want the broader picture, the best coding model roundup and best AI for web development are the natural next reads.
