Last week I ran the same 800-word scene through eleven models. Same prompt, same story bible, same character voice notes. The priciest one cost roughly forty times what the cheapest paid model did. The prose was not forty times better. Squint at the two side by side and you couldn't reliably say which was which.
That gap is the whole argument. Go hunting for the cheapest AI model for writing fiction and the trap is assuming "cheapest" means "free model that writes like an HR announcement." It doesn't. There's a fat middle band on OpenRouter where you pay pennies per chapter and the prose holds its own against models that want your credit card on a monthly plan. That band is where you should live.
Here's the math, then the ranking.
What a chapter of fiction actually costs
Nobody thinks in tokens while drafting, so let's convert to something you can feel.
Rough rule: 1,000 words of output runs about 1,300–1,400 tokens. When you expand a beat, the model also reads your prompt, your recent prose, and whatever the story bible dragged in — call it another 2,000–4,000 tokens of input per call. So one "expand this beat into 1,000 words" request lands around 3,500–5,500 tokens total, most of it on the cheaper input side of the ledger.
Now price that against real rates. A model at $0.14 in / $0.28 out per million tokens — roughly where DeepSeek's mid-tier sits — costs about a tenth of a cent to draft 1,000 words. A full 3,000-word chapter, generated fresh: under half a cent. You'd have to write a hundred chapters to spend fifty cents.
Compare the "premium" shelf. A flagship creative model at $0.50 in / $3.00 out per million runs roughly ten times more per word. Still cheap in absolute terms — a chapter is maybe three cents — but ten times cheap is still ten times cheap. Over a 90,000-word novel with the usual three or four regenerations per scene, that's the difference between real money and a rounding error.
Prices shift constantly, so check the live OpenRouter pricing page before you commit. The numbers below are the July 2026 shape of things, not a promise.
Ranked by whether the prose is any good
I'm ranking by prose quality first, price second — because a free model that writes flat sentences isn't cheap, it's expensive in rewrites. Every one of these runs on your own OpenRouter key, so you pay the model provider directly.
The sweet spot: DeepSeek's mid-tier line
For most fiction, start here. DeepSeek's V3.2 / V4 chat models sit around $0.14/$0.28 per million and punch absurdly above that weight. Long context (the V4 line reaches a million tokens), coherent across a full chapter, and — the part that matters — they don't fight your voice. Feed them a gritty noir sample and they stay gritty. They're strongest when the story leans on logic and structure: mysteries, plot-heavy fantasy, anything where a reader will catch a continuity slip three chapters later.
Weakness: they go a touch plain on the purely lyrical passages. Not bad, just not showy. For most drafting, plain-that-reads-clean beats purple-that-needs-cutting.
The stylist on a budget: Gemini Flash
Gemini's Flash tier (around $0.50 in / $3 out) is what I reach for when a scene needs to sing — deep interiority, an emotional gut-punch, a narrator with a real accent to the voice. On the creative-writing leaderboards it lands shockingly close to the flagship Pro model at a fraction of the cost. Call it 85% of the quality for a sliver of the price, and for line-level prose that 85% is usually the 85% you'd have kept anyway.
It costs more per word than DeepSeek, so I don't run whole first drafts through it. Draft cheap, then send Flash the scenes that carry weight.
Free, with an asterisk
OpenRouter hosts genuinely free models — DeepSeek R1, Llama 3.3 70B, Gemma 3, and others rotate through at zero per-token cost. The catch is rate limits (on the order of 20 requests a minute, a couple hundred a day) and no promise any given free model sticks around next month.
For prose, free Llama 70B and free DeepSeek R1 are legitimately usable — better than most people expect. I've drafted whole scenes on them. If your budget is genuinely zero and you can work around the daily cap, start there. I wrote a fuller breakdown of the best free models for fiction if that's your lane.
The honest catch: free models are where voice-drift bites hardest. They slide toward the tidy, over-explained, everything-wraps-up-neatly rhythm that reads as machine-made. You'll rewrite more. That's what you're paying instead of pennies.
Spending the least without writing worse
The cheapest setup isn't one model. It's a workflow that spends the expensive tokens only where they earn it.
- Draft rough on the cheap tier. Get the beats down with DeepSeek mid-tier or a free model. First drafts are supposed to be ugly; don't pay stylist prices for sentences you'll delete.
- Promote only the keepers. When a scene is structurally right and just needs the prose lifted, run that one through Flash. You're paying the premium rate on maybe a fifth of your words.
- Keep your story bible tight. Every call re-reads context. A bloated bible means paying to feed the model backstory it doesn't need for this scene. Lean bible, fewer input tokens, cheaper every single call.
- Regenerate on purpose. The real cost sink isn't the model, it's hitting Expand eight times because the prompt was mush. Write a sharper beat and you'll get a usable draft in one or two passes.
Do that and a whole novel's drafting cost lands somewhere between "free" and "two coffees." Not a subscription. Not a per-seat license. Just the raw model price, which on the cheap tier is almost nothing.
The part nobody mentions: markup
Here's what makes "cheapest" slippery everywhere else. Plenty of writing apps quote a flat monthly fee and quietly run a cheap model behind the curtain, pocketing the spread. You pay $20 a month for something that costs them forty cents in tokens.
That spread is exactly what I built KudoWrite to kill. You bring your own OpenRouter key, the app sends your prompt straight to the model you picked, and you pay OpenRouter's raw rate — the pricing-page number, no markup, no middleman cut. Draft on the free tier at literal zero, switch to DeepSeek or Flash when a scene deserves it. The app never sees your card and never sees your writing; your manuscript saves to your own Google Drive, not someone's server. And there are no content filters, so genre, dark themes, explicit scenes — whatever you're actually writing — none of it gets refused or sanded down.
If the subscription-versus-tokens math is the thing nagging you, I laid out the full case for bringing your own model instead of paying a subscription — same argument as this one, aimed at the wallet instead of the prose.
So the cheapest AI model for writing fiction isn't one name. It's a cheap workhorse for volume, a stylist for the scenes that carry the book, and an app that doesn't skim either transaction. Pick that combination and you'll spend less on an entire novel than one month of the tools charging a subscription to write it.
Sources: OpenRouter pricing, OpenRouter roleplay & creative writing collection, OpenRouter: open-weight models that matter, June 2026