Free AI Models for Fiction (October 2026): What Is Actually Free on OpenRouter

The free AI models for fiction worth your time on OpenRouter this month are four: Google's Gemma 4 31B, NVIDIA's Nemotron 3 Super and Nemotron 3 Ultra, and MiniMax M3.1 Flash. They cost nothing per token, they are capped at 20 requests a minute and 50 a day on a fresh account, and the best of them writes a usable scene. The rest of OpenRouter's free list is coding agents, a content-safety classifier and a 2.6B model, none of which you want near a manuscript.

Everything below was checked on 19 October 2026 against OpenRouter's own model list and rate-limit docs, and one short beat was run through the four models above on an OpenRouter key. The lineup rotates: the Hermes 405B, Qwen3-Next and Llama 3.3 70B free ids that older lists (including an earlier version of this page) recommended are paid now.

Which free AI models for fiction exist on OpenRouter right now

OpenRouter marks a free endpoint with a :free suffix on the model id. On 19 October 2026 the list held 469 models and 17 of them carried that suffix, all of them below.

Ids and descriptions from openrouter.ai/models, 19 October 2026. The free id is the same page as the paid one, with the variant switched.

Every free model that can write is a reasoning model except Gemma, which matters for the daily cap (more below). And there is no free DeepSeek, no free Kimi, no free Llama and no free Qwen this month. If a guide tells you otherwise, check the date on it.

The free-model rate limits, as OpenRouter documents them

OpenRouter's limits page gives free variants 20 requests per minute. The daily ceiling depends on credits you have ever bought: 50 requests a day on an account that has bought fewer than 10 credits (dollars), 1,000 a day once you have bought at least 10. The FAQ puts it the same way: "low rate limits (50 requests per day total)" until you have put $10 on the account. Those same 50 a day are the ration on a free Janitor AI proxy, where a failed message counts too.

On top of OpenRouter's own gates sits the provider's. Our first call to Gemma 4 31B came back 429 with the message that the model "is temporarily rate-limited upstream", a shared pool at Google AI Studio that every free user on OpenRouter is drawing from. The retry a few minutes later went through. Nemotron 3.5 Super VL returned a 502, then timed out after two minutes. Free means you queue behind everyone else who is also paying nothing.

Fifty requests a day is about fifty beats. For a writer working one passage at a time that is an evening's drafting. It is not a chapter-a-click workflow, and regenerating a beat six times burns six of them.

One beat, four free models

The beat: Nadia hears her landlord's key in the lock while she is still hiding the eviction letter. Each model got the same system prompt (close third person, present tense, Nadia's point of view, 120 to 160 words, prose only) and the same one-line beat, through https://openrouter.ai/api/v1/chat/completions on 19 October 2026, temperature 0.8 (providers disagree on what a good temperature for creative writing is, so we fixed one). The lines below are quoted exactly as returned. The token counts and cost are what OpenRouter reported in the response.

nvidia/nemotron-3-ultra-550b-a55b:free, via OpenRouter, October 2026. 81 prompt tokens, 574 completion tokens (400 of them reasoning), cost $0.00, 10 seconds.

The bolt throws. Heavy. Final.

She straightens, wipes her palms on her thighs. The kitchen smells of last night's onions and the lemon cleaner she'd used this morning, desperate for normalcy. Through the thin wall, Mrs. Kovac's television murmurs a laugh track.

The door swings inward. Mr. Henderson fills the frame, his shoulders broad as a doorframe, a manila folder tucked under one arm. He doesn't smile. He never does.

Verdict: the one to draft on. It wrote a room (the onions, the laugh track through the wall, a fridge magnet from an aquarium trip in the lines before these) and it let a three-word paragraph sit. It also wrote a man who fills a door frame with shoulders "broad as a doorframe", which is the kind of line you cut on the second read. 574 tokens for 160 words of prose, because 400 went on thinking you never see.

minimax/minimax-m3.1-flash:free, via OpenRouter, October 2026. 255 prompt tokens, 4,519 completion tokens (4,319 of them reasoning), cost $0.00, 26 seconds.

No knock. Mr. Varga never knocked.

Her stomach tightened. If he saw the letter, he would know she had received it. If he did not, he would know she was afraid.

"You're still here," he said.

Nadia pressed her palm over the envelope. "I was getting some air."

Verdict: the only one of the four that wrote the scene as a scene, with the landlord in the room and a lie in Nadia's mouth, and the only one that understood the letter is a secret she keeps from him. It ignored the present tense. And it is expensive in the one currency free models charge: our first attempt, with a 1,500-token ceiling, returned nothing at all because every token went on reasoning. Asking OpenRouter to turn reasoning off got a 400, "Reasoning is mandatory for this endpoint and cannot be disabled." Give it room (we allowed 8,000 tokens) and wait.

google/gemma-4-31b-it:free, via OpenRouter, October 2026. 67 prompt tokens, 177 completion tokens (no reasoning), cost $0.00, 6 seconds, on the second attempt after a 429.

The sound vibrates through the floorboards, rattling her teeth. Mr. Henderson is early. She freezes, her heart hammering a frantic rhythm against her ribs, her fingertips still tingling from the friction of the paper. The lock clicks open with a finality that feels like a gavel hitting a bench.

Verdict: the cleanest sentences and the most stock ones. A heart hammering against ribs, a gavel, a "brittle smile" in the line after. It also dressed the set wrong: an eviction notice on "heavy vellum" and a "mahogany side table" for a tenant who can't make rent. Fast, obedient, no reasoning tax, and it held the tense. Good for a tidy pass over a beat you have already shaped. Not where the voice comes from.

nvidia/nemotron-3-super-120b-a12b:free, via OpenRouter, October 2026. 81 prompt tokens, 259 completion tokens (87 reasoning), cost $0.00, 6 seconds.

She presses the eviction letter tighter against her chest, the paper's edge biting into her palm, and slides it behind the stack of unpaid bills on the kitchen counter, hoping the thin veneer of order will mask the dread coiling inside her. The door creaks open a fraction, a sliver of hallway light spilling onto the worn linoleum, and she freezes, breath held, eyes darting to the shadow that looms just beyond the frame.

Verdict: competent and overwritten. Every noun gets an adjective, every feeling gets named ("the dread coiling inside her"), and the whole passage is four long sentences with the same shape. It never let the landlord in. Usable as raw material if you cut half the words. Of the four, this is the one that reads most like a machine.

Two of the four named the landlord Mr. Henderson without being asked. Keep that in mind when a free model's choices feel inspired.

The paid models that cost about a cent a chapter

The honest comparison isn't free versus a $20-a-month writing tool. It's free versus the cheap paid tier on the same key, where the per-token prices are small enough that a whole novel's draft costs less than a coffee. These are the list prices on each model's OpenRouter page, checked 19 October 2026, per million tokens.

What those numbers mean on the page: a 1,000-word passage is roughly 1,300 to 1,400 tokens of output. At DeepSeek V4.1 Flash's $0.08 per million output tokens, the words of a 90,000-word novel (about 120,000 output tokens) come to about one cent, before the input tokens each call reads and before you regenerate anything. Even with the reasoning tokens these models spend, and three or four passes per scene, you are in the tens of cents for a book. OpenRouter's FAQ states it passes through provider pricing "without any markup"; the fee is on buying credits, 5.5% with an $0.80 minimum by card.

The same Nadia beat through the cheapest of them:

deepseek/deepseek-v4.1-flash, via OpenRouter, October 2026. 92 prompt tokens, 934 completion tokens (684 reasoning), cost $0.0001118 as reported by OpenRouter, 7 seconds.

Nadia freezes, the eviction letter half-shoved under the toaster. Its edge juts out, white as a bone. She presses it down with her palm, feels the cheap paper crinkle against her skin. The door swings open. Mr. Alvarez's boots thud on the linoleum, slower than usual, deliberate. He doesn't call her name.

Verdict: a hundredth of a cent, and it is the passage we would keep. It held the tense, it hid the letter somewhere a real person hides things, and in the lines after these it has Nadia turn the tap on and run cold water over her wrists so it looks as if she has been washing up all along. That is the thing the free models mostly didn't do: let her act instead of feel. "White as a bone" is a stock simile. Cut it and the paragraph is done.

So the pattern that spends least without writing worse: draft on a free model while you have requests left, promote the scenes that carry weight to a cheap paid model, and keep the stylist tier (Gemini Flash and up) for the handful of passages that have to sing. For the top of that ladder, the Claude and Gemini picks and what they cost, see the best OpenRouter models for creative writing.

Where the free AI models for fiction fall short

Fifty requests a day, shared upstream pools, and no promise the id exists next month. This session turned up three more.

Reasoning you pay for in silence. Three of the four free writers are reasoning models, and the thinking counts against your output budget: MiniMax spent 4,319 tokens deciding what to write before it wrote 180 words. Set the output ceiling low to save time and you get an empty reply and a spent request.

Instructions slip. MiniMax wrote the whole beat in past tense after being told present. Nemotron Super never got the landlord through the door. None of this is fatal, but each slip is a regeneration, and a regeneration is one of your fifty.

The set dressing is generic. Henderson twice, a hammering heart in three of the four, linoleum in two. A free model's first draft is a draft of the average eviction scene. Your steer is what makes it Nadia's.

And there is no uncensored free model this month. The Dolphin Venice fine-tune that used to be the free pick for dark or explicit fiction is a paid id now, at $0.20 in and $0.90 out. Gemma and Nemotron are instruct models with their makers' guardrails. For horror or explicit romance, budget the cents.

How to write with them

You need an OpenRouter account and a key; what OpenRouter is, for writers walks through that in plain English. Then any app that takes a key can use the free ids. Before you paste a chapter, check whether AI trains on your writing: OpenRouter has a switch for providers that train, and free models have their own.

Or skip the free pool and its 429s. KudoNovel needs no key at all: its built-in AI, KudoAI, runs on a current DeepSeek "flash" model, the cheap paid tier from the table, picked weekly for price and routed only to zero-data-retention hosts that don't train on prompts. There's no content filter of KudoNovel's own, and the model was chosen partly because it rarely refuses, which the free ids above can't promise for horror or explicit romance. Every beat is written from a story bible and a summary of the book that keep up as you write, so the set dressing is your book's, not the average eviction scene. Confirm your email and one free credit is there to write a beat; after that, one-off packs, $3.50 for 100 credits, nothing monthly.

Checked October 2026: the 17 :free ids, the 20-a-minute and 50-a-day limits (1,000 a day after $10 of credits bought), the prices in the table and the five samples above are all from 19 October 2026. OpenRouter rotates its free list without notice. When an id in this guide stops answering, open openrouter.ai/models, filter by price, and pick the largest general-purpose instruct or reasoning model with a :free tag that isn't a coding or safety model.

Take your own beat, the one you're stuck on, and run it through Nemotron Ultra and Gemma 4 31B tonight. Two requests out of fifty, and you'll know in five minutes which of them sounds like your book.