What the free WordPress AI plugin actually costs to run

The canonical WordPress AI plugin costs nothing to install and every feature in it spends money. WordPress core ships no AI provider, so you supply an API key and the meter runs on your account. Every tutorial tells you to paste that key. None of them tells you what happens to the bill when you click Generate on a 500-image media library.

So here is the arithmetic, worked out end to end, with the assumptions written down so you can change them. The headline: for a typical small business site the answer is a few dollars a month, and the expensive feature is not the one people worry about.

What does the WordPress AI plugin charge for?

Nothing directly. The canonical AI plugin is the official plugin that carries WordPress AI features outside the core release cycle, and it is free on the plugin directory. What costs money is the model behind it.

The AI Client is the provider-agnostic PHP API that WordPress 7.0 added so plugins can send prompts through one interface. Core deliberately bundles no providers alongside it. Those arrive as separate plugins for OpenAI, Anthropic and Google, and each one asks for your own API key (Make WordPress Core). Every generation the plugin performs is a request billed to that key.

The feature list grew twice this summer. Version 1.2.0, released July 21, 2026, added bulk content summaries and a comment-reply experiment. Version 1.3.0, released August 18, 2026, added slug generation, a content translation experiment and improved classification (Make WordPress AI). Each of those is another button that spends tokens.

How much does a month of it cost?

Between two and five dollars for a normal small site, at August 2026 list prices. Here is the modelled month:

Feature	Volume	Tokens in	Tokens out	Haiku 4.5	Sonnet 5
Alt text	500 images	683,500	15,000	$0.76	$1.52
Classification	200 posts	400,000	12,000	$0.46	$0.92
Slug generation	200 titles	44,000	3,000	$0.06	$0.12
Bulk summaries	200 posts	400,000	24,000	$0.52	$1.04
Translation	50 posts	85,000	105,000	$0.61	$1.22
Total		1.61M	159,000	$2.41	$4.81

The same month on Claude Opus 5 comes to $12.04, which is the clearest argument in the table for not using a frontier model to write alt text.

Prices used are the published list rates on August 22, 2026: Claude Haiku 4.5 at $1 per million input tokens and $5 per million output, Claude Sonnet 5 at $2 and $10, and Claude Opus 5 at $5 and $25 (Claude pricing).

How was that calculated?

By counting tokens per job and multiplying. A visual token is the unit Claude bills images in, equivalent to one 28x28 pixel patch of the image. The assumptions are the part worth arguing with, so they are all here:

1. Images are 1200x800 pixels. Claude views images in 28x28 pixel patches, so an image costs ceil(width / 28) x ceil(height / 28) visual tokens. That gives 43 x 29, or 1,247 tokens per image (Claude vision docs).
2. Each alt-text call adds 120 prompt tokens and returns 30. Alt text is one short sentence.
3. A post is 1,200 words, about 1,600 tokens. Classification and summaries send the post plus context, so 2,000 tokens in.
4. Classification returns 60 tokens, a summary 120.
5. Translation returns as much as it receives. A 1,700-token post produces roughly 2,100 tokens of translated output.

The documentation gives an independent check on the image line. Anthropic's own worked example puts a 1000x1000 image at 1,296 tokens, or about $1.30 per thousand images on Haiku 4.5. The 500 images above, at a slightly smaller token count, land at $0.76, which is consistent with that.

Two caveats. Providers count image tokens differently, so the alt-text row is specific to Claude models; the text-only rows port more cleanly. And retries are not modelled. A failed generation you click again is a second bill.

Which feature actually costs the most?

Translation, by a wide margin per item. Everything else is a cheap read and a short answer. Translation is a long read and an equally long write, and output tokens cost five times input on every model in the table.

Per unit, the ranking is stable across providers:

Feature	Cost per item, Haiku 4.5	Why
Translation	$0.0122	Long output, billed at the output rate
Alt text	$0.0015	Image tokens dominate, output is tiny
Bulk summary	$0.0026	Full post in, paragraph out
Classification	$0.0023	Full post in, a few labels out
Slug	$0.0003	A title in, a few words out

That reverses the intuition most people bring to this. Alt text feels expensive because images feel heavy, but a 1200x800 image is about 1,247 tokens, which is less than the post it sits in. Translation feels free because it is one click, and it is the priciest thing in the plugin.

For example, translating an entire 200-post archive into one language, rather than the 50 posts modelled above, costs about $2.44 on Haiku 4.5 on its own. That is still small, but it is the line that scales with ambition.

Can I run it cheaper on another provider?

Yes, and the text-only work gets very cheap indeed. Dropping the alt-text row and pricing only the four text features:

Model	Text-only month	List price in / out
gpt-5-nano	$0.10	$0.05 / $0.40 per million
gpt-5-mini	$0.52	$0.25 / $2.00 per million
Gemini 3.5 Flash-Lite	$0.64	$0.30 / $2.50 per million
Claude Haiku 4.5	$1.65	$1.00 / $5.00 per million

OpenAI's published rates put gpt-5-nano at $0.05 per million input tokens and $0.40 per million output (OpenAI pricing), and Google lists Gemini 3.5 Flash-Lite at $0.30 and $2.50 (Gemini API pricing).

Cheaper per token is not automatically cheaper per finished job, though. A model that returns a bad slug or a wrong category costs a second call plus the editor's time to notice, and the cheapest AI API is not the cheapest to run once those retries land in the bill. Test the small model on your own content before committing an archive to it.

So is it worth turning on?

Yes, on cost grounds. Two to five dollars a month is less than any commercial AI plugin licence, and less than fifteen minutes of the time it takes to write 500 alt-text strings by hand.

The decision that actually matters is which model you point it at, and the rule is simple:

1. Use a small model for alt text, slugs and classification. These are short, mechanical outputs where a frontier model buys nothing.
2. Consider a mid-tier model for translation and summaries. Quality shows in longer prose, and these are your expensive calls anyway.
3. Set a spending cap on the provider account. This is the real protection against a bulk action nobody meant to run.
4. Watch for repeat runs. A bulk job re-run over the same archive bills again in full, and Bedrock cache writes ate 85 percent of one AI bill for a related reason: it is rarely the call you planned that hurts.

The number is small. Know it anyway

The useful finding here is not that WordPress AI is expensive. It is that it is cheap, and that nobody publishes the figure, so site owners either avoid the features or turn them on blind.

Run the same arithmetic on your own numbers before a bulk action. Count the items, multiply by the per-item costs above, and you will know within a dollar what a click is about to spend. That is a two-minute check, and it is the difference between a considered decision and a surprise on next month's provider invoice.