Proofreading WordPress posts with AI: what it costs and what to watch

October 22, 2026 · 8 min read

Rules can find double spaces and misplaced commas, but not "teh", "recieve" or "they was". Large language models catch those easily, and with your own API key, proofreading a post costs very little. This guide shows how to estimate the cost, what gets sent, and how to use AI suggestions without letting a model rewrite your content.

Three ways to proofread with AI

1. Copy and paste into a chat app. Free or included in a subscription, but slow, it loses formatting, and the corrected text has to be pasted back by hand — with the risk of overwriting links and blocks.
2. A plugin with its own AI subscription. Convenient, but you pay a monthly fee and your text goes through the vendor's servers.
3. A plugin that uses your own API key. You pay the AI provider directly, at their per-token rates, and the text goes straight from your site to the provider.

What it costs

AI APIs charge per token — roughly ¾ of an English word; text in other alphabets often needs more tokens per word. Proofreading is cheap because the input is your text and the output is only a short list of corrections, not a rewritten article.

A rough estimate for one 1,000-word post: about 1,500 tokens of text plus about 1,000 tokens of instructions in, and a few hundred tokens of suggestions out. Some models also spend tokens on internal reasoning, which is billed as output. With Anthropic's prices at the time of writing:

So even proofreading 500 posts with the most capable model costs in the tens of dollars, and a fraction of that with smaller models. OpenAI models are in a similar range — check the current prices on the OpenAI pricing page and Anthropic pricing page, as they change.

Costs grow when a tool sends the whole post again for every small question, asks the model to rewrite the full text, or re-checks posts that have not changed. Caching answers for unchanged posts and setting a daily request limit keep the bill predictable.

What is sent, and where

• Send only what needs checking. The visible text is enough; markup, links and e-mail addresses do not need to leave your site.
• Know the provider's terms. Both Anthropic and OpenAI state that data sent through their commercial APIs is not used to train models by default; retention periods and exceptions are described in their policies (Anthropic, OpenAI).
• Tell your users. If authors or clients write on the site, mention AI proofreading in your privacy policy, and do not send drafts that contain personal data.
• Protect the key. Store the API key in wp-config.php rather than the database when the tool allows it, and set a spending limit in the provider's console.

Using suggestions safely

• Never apply AI changes automatically. Models occasionally "correct" proper names, product terms, quotes, deliberate style and regional spelling.
• Prefer suggestions over rewrites. A list of small, specific corrections is easier to review than a rewritten paragraph — and cannot silently change meaning elsewhere.
• Tell the model the language. Otherwise British spelling becomes American and French spacing gets "fixed".
• Keep rules for the mechanical stuff. Spaces, punctuation and dashes are faster, free and fully predictable with rules; use AI for what rules cannot see.

AI proofreading in Content Polisher Pro

Content Polisher Pro adds AI proofreading on top of the rule-based checks, designed around the points above:

• Your own Anthropic Claude or OpenAI API key, which can be defined in wp-config.php.
• Runs only when you click Proofread with AI on a post and confirm — never during regular scans.
• Sends the visible text of the post and its enabled ACF fields in one request, without markup, links or e-mail addresses.
• Returns suggestions in the same Before / After format as every other issue — nothing is applied automatically, and every applied fix can be undone.
• Caches answers for unchanged posts and enforces a daily request limit.

See how to set it up.