Which AI writes the best German? Eight language models in a blind test

In short: In our blind test of eight language models and around 270 texts, OpenAI's GPT-6.1 Sol wrote the best German – first place in every rating it took part in. GPT-6 Luna offers the best balance of quality and price: always in the top group, at a fraction of the cost. The writing specialist Hemmingway-1 and the new Mistral Large 4 ended up near the bottom. Hermes 3000 therefore now writes with GPT-6 Luna by default; paid books can additionally use GPT-6.1 Sol. As of 16 October 2026.

You can't tell which language model is behind a writing app from a single sentence. You notice it after forty pages: when the narrator suddenly says "I" although the book is written in the third person. When a character walks down a ramp that hasn't been lowered yet. When on every third page someone runs "their thumb over" something. These mistakes are small, but they cost time in revision – and they decide whether a writing assistant takes work off your hands or creates more of it.

Several new models have been released since our last comparison in the summer, some of them much cheaper. And with Hemmingway-1 there is a model trained explicitly "to write like a person". Reason enough to test thoroughly once more. This workshop report explains how we went about it, what came out – and what we did with it.

The candidates

These are the providers' list prices at the time of the test. Important: many of these models "think" before they write, and that thinking is billed as output. What a model really costs per page only shows up in the measurement – more on that below.

How we tested

Eight tasks, seven of them in German. An opening scene, a dialogue with an unspoken conflict, a description of a place that shows feelings only through perception. Plus three tasks close to how people write in Hermes: continuing a scene from a novel – with characters, the story so far and the last paragraph –, the opening of a non-fiction chapter, and the beginning of a novel chapter of about 900 words with fixed requirements (third-person narrator, past tense, exactly three passengers, a phone call at the end). Long texts matter: grammar and logic errors pile up with length, and short samples hide them. The eighth task was copy-editing – a paragraph with eleven planted mistakes –, plus an English scene as a control.

Rated blind, from several sides. Every model wrote every task several times. Two AI raters from different companies – Anthropic's Claude Opus and OpenAI's GPT-5.5 – received all texts for a task without model names, in an order shuffled separately for each rater, in the role of a publishing editor. They scored language, style and adherence to the brief, had to quote every error verbatim and rank the texts. On top of that we turned the texts into a blind-reading page and had Claude and ChatGPT judge them as free readers.

Why several raters? Because every AI rater has a relative in the field: Claude Opus also judges Claude Haiku, ChatGPT also the GPT models. A result only one rater sees is suspicious. One that all of them see holds up.

And then once more from scratch. We pitted the top group against each other in a second round with fresh texts – three new versions per task that no rater had ever seen. Mistral Large 4, which was released as a preview during our test, joined in that round. In total, around 270 texts were written.

The result

The second round with fresh texts is the most meaningful. Average rank means: position among the six texts of a task, averaged over all tasks (1 = best). Scores for language, style and adherence run from 1 to 10.

In the first round with all models, the previous default GPT-5.4 mini and both Hemmingway variants were clearly behind – sixth to eighth out of eight for both raters.

GPT-6.1 Sol doesn't win narrowly but clearly: 25 of 32 verdicts in the first round, 38 of 48 with fresh texts. In the free blind reading by Claude it won 13 of 16 tasks; ChatGPT, which already knew the model assignment at that point, came to the same result. When continuing a story with book context and in the long chapters – exactly where a writing assistant does the most work – the raters found only minor issues in its texts, no grammar mistakes and no breaks in logic. Claude Opus summed up what Sol does differently: it almost never explains feelings. The meaning lies in a gesture or an object.

Als Kind hatte sie dort mit der Zirkelspitze Löcher gestochen, während ihre Mutter am Herd stand. Sieben, dann eine Lücke, dann drei. Ihr Daumen fand die Lücke sofort.

(As a child she had pricked holes there with the point of a compass while her mother stood at the stove. Seven, then a gap, then three. Her thumb found the gap at once.)

Jetzt hob er den Koffer über den Spalt zwischen Anleger und Fähre. Sie hielt die Luft an, bis beide Räder auf dem Deck standen.

(Now he lifted the suitcase over the gap between pier and ferry. She held her breath until both wheels were on the deck.)

GPT-6 Luna is the model all raters agreed on: in the top group in every rating, never weak, almost always first when Sol isn't in the field – at a cost of less than a tenth of a cent per standard page.

Claude Haiku 5.5 writes strong short, atmospheric scenes – in Claude's blind reading it won both opening scenes – but made occasional grammar mistakes in long German texts.

Gemini 4 Argon split the raters like no other model. In copy-editing it was unbeatable and won almost every verdict. For literary writing, Claude Opus saw it in front in one round, while ChatGPT put it last – "technically clean, but smooth". With fresh texts it came in behind GPT-6.1 Sol and GPT-6 Luna.

Mistral Large 4, the French provider's new flagship, was only released as a preview during our test. We added it in the second round. It came last, without winning a single verdict. The raters criticised breaks in logic – "three cups, like old times" with only two people at the table –, clichés and sentences that explain instead of narrate. It also thinks for a long time: about a minute per text on average. Since it is a preview version, we will take another look at the final release.

GPT-5.4 mini, our previous default, landed in the bottom half of every rating. It is fast, but every newer model in the field writes better – most of them for less. In the long chapter it missed the brief: once more than the required three passengers turned up, another time an unknown number called at the end instead of the son.

The writing specialist: Hemmingway-1

We were particularly curious about Hemmingway-1. The model is based on an open 27-billion-parameter model and, according to its makers, was trained specifically "to write like a person". We tested it in two settings: with full thinking time (the default) and with a shortened one.

In our test it did not convince. Both variants ranked near the bottom with the raters who looked at craft and adherence to the brief, in German as well as in English. In the long chapter, each variant once wrote in the first person although a third-person narrator was required, and switched from past to present tense mid-text. There were also ungrammatical sentences. With its default setting it took 89 seconds per text on average, once almost four minutes for a chapter.

To be fair: individual sentences were strong. "Die Bank hat noch die Form von zwei Menschen." (The bench still has the shape of two people.) is an opening no other model found so concisely. And when ChatGPT rated the texts purely for literary originality, Hemmingway came third. The model has outliers at the top – but a writing assistant above all has to be reliable. A text that ignores your book's point of view is useless to you, however good it sounds.

What a model really costs

The list price says little. What counts is the price per written page – including the invisible thinking. Measured in our test, per standard page (1,800 characters):

Calculated with the list prices above and the tokens measured in our test; waiting time as the median.

Two observations: Hemmingway-1 and Mistral Large 4 are not expensive per token, but they think for so long that per page they are among the most expensive models in the field. And GPT-6.1 Sol costs about sixteen times as much per page as GPT-6 Luna – at less than one cent per page, that is still little for a single book.

A habit almost all of them share

Reading the texts, we noticed a habit almost all models share: their characters keep running their thumb over something – a table edge, a timetable, a coffee cup, a screen. In the fresh round, one in five German texts contained a thumb, with GPT-6 Luna more than one in three. On its own it is a nice, concrete detail. Spread over a whole book it becomes a tic.

That is a useful tip for anyone writing with AI: when revising, watch for recurring gestures and images. A model that convinces in one chapter often repeats itself across twenty chapters more than you notice on first reading.

What we are doing with it

• GPT-6 Luna is now the default model for writing in Hermes 3000. It writes better than our previous default and costs less.
• GPT-6.1 Sol is available for paid books – for everyone who has bought a book or uses a membership. Choose it in your book's AI settings. It wrote the best German in our test, especially when continuing a story and in long chapters.
• You can try Sol for free in the model arena. There, up to four models continue the same sentence side by side, and you decide what you like.
• All other models remain selectable – Claude Opus 4.5, which costs about as much per page as Sol, likewise for paid books. Taste in writing is not negotiable, and some authors like precisely the idiosyncratic tone of a particular model.

Limits of this test

We want to be open about a few things so nobody reads more into it than is there. The ratings come from AI models, not from human editors; that is why we used several raters from different companies and had the top group compete again with fresh texts. The tasks were short to medium-length and without a complete manuscript – in real use the model gets much more context. Prices change quickly; Gemini 4 Argon, for example, is expected to cost twice as much from December. And taste remains taste: a model that ranks low in our test may be exactly right for your style.

Frequently asked questions

Which AI writes the best German?

In our blind test of eight models in October 2026, OpenAI's GPT-6.1 Sol wrote the best German – first place in every rating it took part in. GPT-6 Luna had the best balance of quality and price.

Is Hemmingway-1 good for writing books in German?

Not in our test: it ranked near the bottom for craft and adherence to the brief, switched to the first person and present tense unprompted, and took 89 seconds per text on average with its default setting. Individual sentences were strong, though, and in a purely literary rating it came third.

What does writing with AI cost per page?

In our test, a standard page (1,800 characters) cost about 0.04 cents with GPT-6 Luna and about 0.6 cents with GPT-6.1 Sol. What matters is the price per page, not per token, because many models "think" invisibly before writing and that is billed too.

Can I choose the AI model in Hermes 3000?

Yes. In each book's AI settings you choose the default model; GPT-6 Luna is preselected, GPT-6.1 Sol is available for paid books. In the model arena you can try several models side by side – including Sol.

Try it yourself

You'll find the model arena in every book under ⋮ → AI settings. Enter a sentence from your manuscript, choose up to four models – including GPT-6.1 Sol – and read the versions side by side. You'll know which model suits your style after two paragraphs. Hermes 3000 is free up to 50 standard pages: Try it now.