How Much Does It Cost to Rewrite an Article With AI? Price a Whole Backlog, Model by Model

How much does it cost to rewrite an article with AI? For a 1,200-word post refreshed in place, our estimate runs from $0.09 on Gemini 3.8 Flash to $0.85 on GPT-5.6 Terra at the Balanced level, web research and fact-check included, and from $0.02 to $0.14 at Minimal, which skips the research step, billed by your AI provider on your own API key. The calculator below prices one post or a whole backlog, model by model, with the tokens behind each figure.

Cost calculator: price one post or your whole backlog

Set how many posts you want to refresh, their average length, the level of change and the model your site will use. Every figure below is the plugin's own estimate for those inputs, the one it shows before a refresh you confirm from the editor, the posts list or Content health.

How the estimate is built, and what it bounds

The estimate turns your post's word count into tokens, multiplies them by a factor measured for each level on real refreshes, and adds every web search the level is allowed to run, at your provider's published rates.

1,000 words are counted as 1,400 tokens. The search counts are the caps of the pipeline's web steps, not averages.

That 1.4 is on the high side of Anthropic's rule of thumb, "1 token is approximately 4 characters or 0.75 words in English", or 1.33 tokens per word. The Minimal and Balanced factors were measured on 2026-07-11, Balanced on a Claude Sonnet 5 run, Minimal on Claude and Gemini runs; the Google column on 2026-08-21, on Gemini 3.8 Flash. Strong and Maximal are extrapolated from Balanced. The rates are those of Anthropic, OpenAI and Google, read 2026-09-15; the formula is in the docs.

A Minimal refresh measured on 2026-09-05 used 7.75 input tokens per token of article against the 10.9 assumed, and the search count is a cap the Anthropic provider cannot reach, one step request running at most one search. The plugin shows it before a refresh you confirm from the editor, the posts list or Content health.

Why rewriting an article costs more than generating one

A generator pays mostly for output; a refresh pays mostly for input, because it reads your article at every step and, on Anthropic and OpenAI, pays for the search results it reads as input tokens.

Analyze, fact-check, rewrite and link each read the article, enrich reads it from Balanced up, verify only the passages the rewrite added. That reading is the prompt overhead behind the table above: at Balanced, 142.7 input tokens per token of article on Anthropic and OpenAI, 10.9 on Google.

Anthropic's pricing page, read 2026-09-15, says web search results "are counted as input tokens"; OpenAI's says "Search content tokens" are "billed at model rates"; Google bills a grounded search per request and does not put the fetched pages into the billed prompt. Hence the table's two input columns.

What a refresh pays for that a generation does not:

• the original, read at each step;
• the search results, read as input tokens on two of the three providers;
• a second check of any figures the rewrite added.

What each level changes in the text is in the guide to rewrite old posts with AI.

Pricing a backlog: batches, free searches, and the second pass a year later

A backlog is priced post by post, the way the plugin's batch screen does it: the same estimate for each post, added up, with the web searches counted against your provider's monthly allowance.

Content health, in the free plugin, queues up to 50 posts per batch, 20 to a page, and shows the estimated total, searches included, before queuing. One batch cannot reach Google's allowance: 50 posts at Maximal's 23 searches make 1,150, inside the 5,000 free ones a month; several hundred posts refreshed within a month can, and the calculator says so.

Twelve months on, a second pass at Minimal skips research and enrichment and runs the analysis, the fact-check, the rewrite step's typo pass, the second check of new figures and the links: the maintenance pass, the calculator's "Twelve months later" line, at today's catalogue rates.

Each run lands in the journal with its tokens, searches and cost; the CSV export carries the same columns. That is where the estimate meets the bill. Which posts deserve the spend comes first: refresh old WordPress posts. The order of operations before you spend anything is in the refresh checklist.

What the figure does not include

The estimate is your provider's bill for the tokens and searches of a refresh, and nothing else. Unstale adds nothing to it.

1. The Unstale license, flat and separate, on the pricing page.
2. Your own time reading the diff before it goes live: the two optional fields above count it only if you type a number; the plugin does not time it.
3. A step that fails, or a rewrite the plugin refuses as too short, is still billed by the provider, and the journal records it.
4. The journal's searches line: none of the three provider plugins reports a search count back to WordPress today, so the journal marks it as not reported and the logged cost leaves the searches out, where the estimate counts them.
5. Rate changes: the calculator uses the catalogue the plugin refreshes once a day, dated 2026-08-22. Google's pricing page, read 2026-09-15, dates its Gemini 3.8 Flash rate to the end of 2026 and lists a higher one from January 2027; Anthropic made the Claude Sonnet 5 launch rate permanent.

Questions about the cost of rewriting with AI

Is it cheaper to rewrite an old article with AI than to generate a new one?

Per word, no: a refresh reads the existing text at every step and, on Anthropic and OpenAI, pays for search results as input tokens, so its input bill is many times the article's length; a generation pays mostly for output. It keeps the URL, and every link already pointing at it.

What is the best AI for rewriting an article, on cost?

On cost, Gemini 3.8 Flash is the cheapest model in the plugin's catalogue: $0.75 against $2 per million input tokens on Claude Sonnet 5, the plugin's default, plus 5,000 free grounded searches a month; GPT-5.6 Terra is the dearest. Which one writes well is what the review queue's diff shows, not a price list.

Does the estimate include web searches?

Yes, at each level's cap: 7 at Minimal, 13 at Balanced, 18 at Strong, 23 at Maximal, priced at catalogue rates: $10 per 1,000 on Anthropic and OpenAI, $14 per 1,000 on Google past 5,000 free a month. The cost logged afterwards does not, the provider plugins reporting no search count back yet.

Will the estimate match my provider's invoice?

It is set on the high side: its token factors were calibrated on real runs, it counts every search the level allows, and a Minimal run in September 2026 used under eight input tokens per token of article against the eleven assumed. The journal records the tokens and cost actually spent.