Sudowrite Adds Kimi K3, the Model Ranked Second for Creative Writing

Sudowrite, one of the better known AI writing tools aimed at novelists, added Kimi K3 to its Write, Draft and Plugins features in an update posted to the company's changelog on August 18, 2026. The same release fixed a long standing problem with Undo and cleaned up a batch of bugs. For authors who pick a writing app based on which models it can drive, this is the kind of update that actually changes the output on the page.

Which model Sudowrite added

Kimi K3 comes from Moonshot AI and was released on July 16, 2026, with the full weights published later that month under Moonshot's own Kimi K3 licence. That licence makes it open weight rather than open source in the formal sense, but the practical effect is that anyone can host it. The model runs 2.8 trillion parameters with a context window of roughly one million tokens, and API pricing sits at about $3 per million input tokens and $15 per million output tokens, according to Moonshot's model documentation and hosting write ups from Northflank and nexos.ai.

The million token context is the number worth pausing on if you write long. It means a model can in principle hold an entire novel manuscript in a single session rather than working from a summary of what came before. Whether a tool actually uses that headroom depends on how the tool is built, not just on the model, so treat it as a ceiling rather than a promise.

Why it matters for fiction specifically

Sudowrite's own note says K3 sits second on the creative writing leaderboard run by EQ-Bench, behind Claude Opus 5. The company describes the model as unusually good at register, meaning that if you ask for hardboiled or lush prose, the sentence rhythm shifts rather than just the adjectives. That is a fair description of what separates a usable draft from one you rewrite line by line.

It is worth knowing how that leaderboard works before you read too much into the ranking. EQ-Bench's creative writing benchmark has the tested model answer 32 prompts across three runs, then scores each result with a judge model against a rubric and derives an Elo from pairwise comparisons. The judge is itself a Claude model, and the benchmark's own documentation says plainly that creative quality is subjective and the judge's assessment may differ from human preferences. Rankings also move as new models land. So a second place finish is a signal that K3 is competitive at prose, not proof that it will suit your book.

The Undo fix is the quieter win

The other change in the release solves a problem that will be familiar if you use chat based editing. Previously, if Chat had edited your manuscript earlier in a writing session, pressing Ctrl or Cmd plus Z later could roll the whole document back past that edit instead of undoing only your most recent manual change. Undo now reverses changes in the order they happened. Sudowrite also made the comment column resizable, reworded the Draft input limit warning, and fixed bugs across highlights, word counts, pasted text, long requests, Brainstorm, Worldbuilding and mobile dictation.

None of that is headline material, but losing an hour of manual edits to a mistimed Undo is the sort of thing that makes writers abandon a tool. Reliability work counts.

What this means for you as an author

The practical takeaway is not that you should switch tools. It is that the model layer underneath your writing app is now something worth checking. A year ago most AI writing tools ran one model and you took what you were given. Now the same app can offer several, and the one that suits a tense action scene is often not the one that suits a reflective chapter opening.

If your tool lets you choose, run the same scene through two models and compare. Pay attention to sentence length, to how quickly the prose drifts back toward generic phrasing, and to whether a character's voice survives a full page. Those three things tell you more in ten minutes than any leaderboard will. Our roundup of the best AI creative writing tools covers what to look for when you are comparing apps rather than models, and our earlier piece on what Anthropic's newest models mean for authors walks through the same exercise for Claude.

Keep the disclosure side in mind too. Amazon KDP requires disclosure of AI generated text, and swapping models does not change what you owe a retailer or a publisher. Whichever model drafts your first pass, keep a record of your own revision work so you can show where the human editing happened.

PageWriter Studio takes the same approach from the other direction. It builds chapters from a style profile trained on your own writing, so the voice stays yours regardless of which model is doing the heavy lifting underneath. If you want to see how that works on a book you are already drafting, you can start a free trial.