August 31, 2026 · 4 min

How Restyler 1 compares to Haiku, Flash Lite and Luna

Everyone assumes AI writing is getting harder to spot. We measured it, across 2,187 prompts, and the opposite is true.

The test is a classifier two-sample test: train a fresh discriminator to separate a system’s output from genuine human writing on the same prompts, and report how often it succeeds. 0.500 means indistinguishable from a person. 1.000 means trivially obvious. Topics are matched by construction, so the classifier has to use style rather than subject.

How the finished text compares

We compare against the fast, cheap tier of each lab: Haiku, Flash Lite, Luna. That is the honest peer group for a model costing $0.15 per million tokens. Here is how machine-written the text a reader actually ends up with reads:

Three different labs, three different families, all within 0.005 of each other, and a full 0.16 worse than the 2024 models they replaced. Better writing, more obviously machine writing.

To be straight about what that table is: Restyler 1 does not write from scratch. It finishes what your model started. But the text a reader ends up with is the thing that matters, and on that measure it lands roughly a quarter of the scale closer to human than anything the fast tier produces on its own.

It is not markdown, either. Strip every heading and bold run and the frontier gap only closes by about 0.05. The rest is prose style: the rhythms, the hedges, the tidy three-part lists. Length is not the driver either. A length-only classifier sits near 0.53–0.62 throughout.

Point it at any model’s output

Restyler 1 is not tied to whatever generated the text. Claude, Gemini, GPT, something you fine-tuned yourself: it takes the response and makes it read like a person wrote it. Same endpoint, same price, whatever is upstream.

Look at the last row. Gemini was held out of training entirely. The model has never seen a single example of how Gemini writes, and it still gets the largest improvement of the three.

That matters more than any single score. You cannot regenerate a training set every time a lab ships a new model, and a rewriter that only worked on the families it had memorised would be obsolete within a quarter. This one transfers, which is what makes it safe to put at the end of a pipeline and leave there.

It keeps what you wrote

Sounding human is the easy half. Anything will sound more human if you let it delete the inconvenient parts, which is exactly what happens when you ask a large model to “make this sound less like AI” and it hands back a tidy summary of two thirds of your page.

On the shipped block-level model, measured on held-out documents:

• 0% invented voice. It never starts speaking as your company. Frontier models do this on 54% of paragraphs.
• 0% repetition loops.
• 64% of rewrites land within 20% of the source length, so a paragraph comes back a paragraph.
• 73% of facts kept on average, and every fact in 41% of blocks.

Size and price

A model built for one job does not need to be large, and not being large is what makes the price possible: $0.03 per million input tokens, $0.15 per million output. A thousand-word article costs about $0.0002 to restyle.

It does not need to be enormous. It needs to have seen that job a great many times.

Try it on something you wrote

The playground runs the same model the API does, with no account and no card. When you want it in your own code it is four lines and an OpenAI-compatible endpoint.

Frontier comparison: 2,187 prompts, topics matched by construction so the classifier must use style rather than subject. Restyler figures are the document-scale evaluation of the Restyler line; the preservation figures are the shipped block-level model on held-out documents. Full numbers and limits on the model page.