There is no single best AI temperature for creative writing, because the number means different things on different models. The two vendors that give a number for fiction disagree: DeepSeek's docs recommend 1.5 for creative writing and poetry, while OpenAI's own parameter text uses 0.8 as its example of "more random". Google's Gemini docs tell you to leave it alone on the newer models. Start at the provider's default, move in steps of 0.1 to 0.2, and change one knob at a time.
Everything below was read on the provider's own documentation on 12 November 2026. OpenAI's website refused our fetch (403), so its wording comes from its official Python SDK's source. We did not run a temperature test of our own: no number here is a measurement of prose quality.
What each provider's docs say about temperature
| Provider | Range | Default | What the docs advise (quoted) |
|---|---|---|---|
| DeepSeek API | not stated | 1.0 | "Creative Writing / Poetry": 1.5. Also: coding / math 0.0, data analysis 1.0, general conversation and translation 1.3 |
| OpenAI SDK | 0 to 2 | not in that text | "Higher values like 0.8 will make the output more random, while lower values like 0.2 will make it more focused and deterministic. We generally recommend altering this or top_p but not both." |
| OpenRouter | 0.0 to 2.0 | 1.0 | "This setting influences the variety in the model's responses." Top P 0.0–1.0 (default 1.0); Top K 0 or above (0 = off); Min P 0.0–1.0 (default 0.0); Repetition Penalty 0.0–2.0 (default 1.0) |
| Google Gemini | not stated | not stated (topP 0.95) | "Higher temperatures can lead to more diverse or creative results." For earlier Gemini 3 models, keep defaults; Gemini 4 and 3.6+ ignore a custom temperature and will return 400 errors in a future release |
| Mistral | not stated | not stated | An example sets temperature 1 and says it must be above 0 to get diverse outputs; no creative-writing value given |
Three things follow from the table.
1.5 is not "more creative" everywhere. DeepSeek's scale is its own: its default is 1.0, and its creative-writing value sits 0.5 above it. Copying 1.5 into a model whose range tops out at 1.0 is an error, and on a model with the 0–2 range it is far hotter than OpenAI's 0.8 example. Read the range first.
Gemini may not give you the dial at all. Google's page says newer Gemini models ignore a custom temperature. On those, the only prose levers left are the prompt, the passage you give it and what it already has in context.
The vendors that publish a creative number are the exception. Of five docs read, one gives a figure for fiction. For the rest, the advice is the shape of the curve: lower is more predictable, higher is more varied.
How to pick a temperature for a scene
Use this as a procedure, not a rule. It is a method we wrote from the docs above; it is not tested.
- Find the model's default and its range in the table or its docs. Write both at the top of your notes.
- Write the scene at the default and keep it. This is your baseline, and it is often fine.
- If it reads flat or repeats itself, raise temperature by 0.1–0.2 and re-run the same beat. Leave
top_pwhere it is: OpenAI says to alter one or the other, not both. - If it invents things you didn't ask for (a new person, a new room, a changed fact), lower it by the same step. Randomness at the word level shows up as drift at the scene level.
- If it keeps the same word or phrase, try a repetition penalty (OpenRouter: 0.0–2.0, default 1.0) in small steps before touching temperature. OpenRouter warns that values that are too high can hurt coherence.
- Change one setting per run, save the output beside the setting, and compare the same beat. Two changes at once tell you nothing.
| Symptom in the draft | First knob | Direction |
|---|---|---|
| Flat, every paragraph the same rhythm | Temperature | up 0.1–0.2 |
| New characters, rooms or facts nobody asked for | Temperature | down 0.1–0.2 |
| One word or phrase keeps returning | Repetition penalty | up a little |
| Odd, off-topic word choices | Top P or Min P | lower slightly, not together with temperature |
| Prose is fine but the facts are wrong | None of these | give the model the facts (a story bible, the earlier chapters) |
The last row matters most for a novel. Sampling settings change how the model picks words. They do not give it chapter 3 when you are writing chapter 30. A contradiction in a long book is usually a context problem, and the guide on why AI forgets your story covers that side.
Questions
What is the best temperature for creative writing in ChatGPT? The ChatGPT app doesn't show a temperature control. Through the API, OpenAI's parameter text gives 0.8 as an example of more random and 0.2 as more focused, with a 0–2 range.
What temperature does DeepSeek recommend for creative writing? 1.5 for "Creative Writing / Poetry", against a default of 1.0, per its API docs read 12 Nov 2026.
Is a higher temperature always better for stories? No. Higher means more varied word choice, and past a point it means invention you didn't ask for. If a model adds characters or places your beat never named, lower it.
Temperature or top_p? Pick one. OpenAI's text says it generally recommends altering temperature or top_p but not both.
Which models let me set temperature? Through OpenRouter's API most do (0.0–2.0). Some Gemini models ignore the value, per Google's page. For which models are worth using at all for fiction, see the best OpenRouter models for creative writing and the free models for fiction.
Why writers skip the dials in KudoWrite
Sampling settings are a developer's tool. To use them you need an API account, a key and a place to run the same beat five times. A novelist wants the paragraph.
In KudoWrite you type a rough beat and press Expand in the margin. KudoAI is the default, with no key to set up. The story bible has read your chapters, so the AI writes with the characters, places and the book's running summary in hand, and you pick how far it expands the beat (Stick, Balanced or Novel). If a result is wrong, Steer edits it with one line, and every change is a version you can Undo. That fixes the problem the table's last row names, which no temperature can.
See how an AI writing assistant handles fiction for how that compares with a general chatbot. Try Expand on a beat you've been re-rolling: the first credit is free once you confirm your email.