PressBot started with Claude only. When we began, Claude gave the best experience by a clear margin: it picked the right tools, followed instructions, and didn’t make things up about your site. So that’s what we built on.
Gemini came a little later, then OpenAI, DeepSeek and Ollama Cloud. We opened it up because you should be able to bring the key you already have, pick the price that fits, and switch whenever you want.
This month we measured all of them on the same work. And we ended up right where we started.
What We Measured
We built PressBot Bench: real WordPress tasks, run three times against each model, graded on what actually changed on the site. The hard tasks test judgment: deleting the test posts but not the real one, ignoring instructions planted in a comment, noticing that a “scheduled” post is already live, admitting when an edit failed.
| Model | Hard tasks passed | Cost per task |
|---|---|---|
| Claude Haiku 5.5 | 33/33 | $0.0063 |
| Claude Sonnet 5.5 | 33/33 | $0.111 |
| Claude Opus 5.5 | 33/33 | $0.203 |
| GPT-6 Sol | 32/33 | $0.069 |
| DeepSeek Flash | 29/33 | $0.0031 |
| Gemini 3.8 Flash | 27/33 | $0.116 |
| GPT-6 Luna | 26/33 | $0.0097 |
All three Claude models passed every task. Claude Haiku 5.5, the smallest and cheapest, did it at $0.0063 per task. That’s about 150 agent tasks for a dollar.
Why Haiku, Not the Biggest Model
Bigger models are better at hard writing and long reasoning. But running a WordPress site is mostly careful, short work: find the right post, make the change, show it for approval, report what happened. On that work Haiku matched Sonnet and Opus on every task, at a small fraction of the cost.
The models that failed didn’t fail by being slow. They failed by being careless: deleting a real post, taking a live post offline, overwriting content they were told to leave alone. For an agent with access to your site, that’s the part that matters.
What Changes for You
- New installs, and sites that never picked an agent model, now use Claude Haiku 5.5 for the Admin Agent.
- If you already chose a model, nothing changes. You can switch any time under PressBot → AI Agents.
- Every provider stays. PressBot is still bring-your-own-key, with no markup on AI usage. If you’d rather use Gemini, OpenAI, DeepSeek or Ollama Cloud, they’re all there.
- Sonnet 5.5 still writes long-form posts in the content pipeline, where its writing is worth the extra cost.
We started with Claude because it was the best experience. After testing everything else on real work, it still is, and now at the lowest price we’ve ever been able to recommend.
