Best Gemini Model for Roleplay (2026): 3.8 Flash, 3.6 Flash or 3.1 Pro?
The best Gemini model for roleplay depends on where you play. In the Gemini app, your plan matters more than your model: it decides whether the game can see 32,000 tokens of your story or 1 million. Through the API — in SillyTavern, Janitor AI or any other frontend — Gemini 3.8 Flash is the default, and Gemini 3.1 Pro is the step up for players who prefer how it writes.
Google’s lineup has changed a lot this year. Four Flash models came out between May and September 2026. In the same months there was no new Pro model. And on 18 September, Google closed the Gemini 2.5 models to new users. If a guide, a Reddit thread or a preset tells you to use “Gemini 2.5 Pro”, it was written for a list that no longer exists.
Everything below was checked on Google’s model list, price list, deprecation table, change log, thinking and safety guides, its Gemini plans page and its Gemini Apps help page, and on OpenRouter’s model list, on 1 October 2026.
A note on method. We build games for Gemini — our Gemini RPG guide explains how — but we have not run a controlled, model-by-model roleplay test of these models. Gemini 3.8 Flash is four weeks old. The advice here comes from what Google publishes about each model, what each one costs over a real campaign, and how a campaign actually uses a model. Where Google describes a model, we quote Google, and we say so.
For the wider question of Gemini against Claude, ChatGPT, Grok and DeepSeek, see the best LLM for roleplay. This page only answers which Gemini.
First: The App or the API?
There are two ways to roleplay with Gemini, and the choice of model works differently in each.
The Gemini app (gemini.google.com and the phone apps). You pick from a short menu, you save your game as a Gem, and your subscription sets the limits. You never see a model list like the one below.
The Gemini API. This is what a frontend such as SillyTavern or Janitor AI connects to, with a key from Google AI Studio or through a router such as OpenRouter. Here you choose the exact model, and you pay per token or use the free tier.
If you play in the app, read the next section and skip the model table. If you play through a frontend, the model table is for you.
In the Gemini App: Your Plan Decides How Much the Game Remembers
This is the most useful fact on the page for app players, and it is not about models at all.
Google’s help page for Gemini Apps lists the context window — how much text the model can read at once — by plan:
| Plan | Price (US) | Context window | Usage limits |
|---|---|---|---|
| Free | $0 | 32k tokens | Standard |
| Google AI Plus | $4.99 a month | 128k tokens | 2× the free limits |
| Google AI Pro | $19.99 a month | 1 million tokens | 4× the free limits |
| Google AI Ultra | $99.99 or $199.99 a month | 1 million tokens | 5× or 20× the Pro limits |
Prices are from Google’s US plans page; they differ by country.
Google’s own measure is that 1 million tokens is “up to 1,500 pages of text”. By the same measure, 32k tokens is about 48 pages. That has to hold everything: your Gem’s instructions, your world file, and the story so far. A campaign fills 48 pages quickly. After that, the model can no longer see the beginning, and Google’s help page describes what follows: responses that “miss connections or details”.
This is the same failure we describe in why AI campaigns fall apart at turn 50. On the free plan it simply arrives sooner.
So for a long campaign in the Gemini app, the thing to buy is the context window, not a smarter model. The step from Free to Plus multiplies the window by four for $4.99. The step to Pro gives you the full 1 million.
Which models the app gives you
Google’s plans page says the free plan includes “Access to 3.6 Flash” and “Varying access to 3.1 Pro”. Its help page shows all plans, free included, with access to three models — Flash-Lite, Flash and Pro — and describes them like this:
- Flash-Lite is “an efficient workhorse model designed for speed”.
- Flash “balances speed and reasoning”.
- Pro is “our most advanced model”, and its replies “will take longer”.
Two details matter for roleplay.
The app is not on the newest Flash. The API’s newest model is Gemini 3.8 Flash. The app’s plans page names 3.6 Flash. You cannot choose 3.8 Flash in the app.
Limits are not a message count. Google says the limits “factor in the complexity of your prompt, the models and features you use, and the length of your chat”, refresh every five hours, and stop at a weekly limit. It adds: “More advanced models and higher thinking levels consume more of your usage.” A long roleplay on Pro uses your allowance faster than the same roleplay on Flash. If you hit the limit on a paid plan, Google says you can continue on Flash-Lite.
The practical reading for app players:
- Free is enough to try a game. Gems are on every plan, including free. Our Gemini RPG prompts are ready to paste into one, and the best Gemini RPG games lists games already built for it.
- Use Flash as your everyday model, and switch to Pro for the scenes that matter. That makes the allowance last.
- For a campaign that runs for weeks, Google AI Pro is the real upgrade, because of the 1M window.
The Current Gemini Models on the API
| Model | Released | Price per million tokens (input / output) | Free API tier? | Default thinking | Google’s label |
|---|---|---|---|---|---|
| Gemini 3.8 Flash | 2 Sep 2026 | $0.75 / $3.75 until 31 Dec 2026, then $1.50 / $7.50 | Yes | Medium | ”Our most intelligent Flash model” |
| Gemini 3.7 Flash | 13 Aug 2026 | $0.75 / $3.75 until 31 Dec 2026, then $1.50 / $7.50 | Yes | Medium | ”Previous-generation Flash model” |
| Gemini 3.6 Flash | 21 Jul 2026 | $0.75 / $3.75 until 31 Dec 2026, then $1.50 / $7.50 | Yes | Medium | ”Previous-generation Flash model” |
| Gemini 3.5 Flash | 19 May 2026 | $1.50 / $9.00 | Yes | Medium | ”Our legacy Flash model” |
| Gemini 3.5 Flash-Lite | 21 Jul 2026 | $0.30 / $2.50 | Yes | Minimal | ”Our fastest, most cost-effective 3.5 model” |
| Gemini 3.1 Flash-Lite | 7 May 2026 | $0.25 / $1.50 | Yes | Minimal | Shutdown no sooner than 7 May 2027 |
| Gemini 3.1 Pro (preview) | 19 Feb 2026 | $2 / $12 up to 200k tokens of prompt; $4 / $18 above | No | High | ”Advanced intelligence, complex problem-solving” |
Release dates are from Google’s deprecation table. Every model in the table accepts about 1 million tokens of input and writes up to about 65,000 tokens of output.
Three things in that table are unusual.
The only Pro model is the oldest model in the list. Gemini 3.1 Pro came out in February and is still labelled a preview. Every Flash model beside it is newer.
The newest Flash costs half as much as the Flash from May. Gemini 3.5 Flash is $1.50 / $9. The three models that replaced it are $0.75 / $3.75. Google’s change log says 3.6 Flash arrived “at a lower price point than 3.5 Flash”.
That low price has an end date. Google calls it an introductory rate. On 1 January 2027, all three newer Flash models double to $1.50 / $7.50.
Which Gemini Model to Pick
Gemini 3.8 Flash — the default
It is the newest model, it has the full 1M window, it has a free tier, and it costs the same as the two Flash models before it. When the newest model is not more expensive, start there.
Know what it was built for, though. Google describes 3.8 Flash as “engineered for long-horizon software engineering, autonomous agents, and complex enterprise workflows”. Nothing in that sentence is about writing a scene. And Google adds a warning of its own:
“Gemini 3.8 Flash can use more tokens on longer running and complex tasks, by design.”
It also tells you what to do about it: “For everyday tasks, you can lower the reasoning effort to reduce token consumption.” A roleplay reply is an everyday task. Set thinking to low. Google’s own description of the low level is that it “reduces time-to-answer for latency-critical tasks like … real-time chat”.
Gemini 3.6 Flash — when you want the fastest replies
Same price as 3.8 Flash, and one feature 3.8 Flash does not have: the minimal thinking level. On 3.8 Flash, Google says minimal “is not supported and returns an error”. On 3.6 Flash it works, and Google describes minimal as the setting that “minimizes latency for chat”.
Google’s change log also says 3.6 Flash was tuned to fix “developer feedback around output verbosity” — in plain words, the model before it wrote too much. For quick back-and-forth with short replies, that is the right direction.
It is also the Flash model Google put in the Gemini app. If you already like how the app’s Flash plays, this is the same model number on the API.
Pick 3.6 Flash if you play fast, chatty scenes and care more about speed than depth.
Gemini 3.1 Pro — for players who prefer its writing
Gemini 3.1 Pro is the model our best LLM for roleplay guide discusses for long, lore-heavy campaigns, and it is the model behind “Pro” in the Gemini app. Some roleplayers prefer its prose to Flash.
On the API it has three costs:
- No free tier. Every Flash model has one. Pro does not.
- About three times the price: $2 / $12 against $0.75 / $3.75.
- A second price above 200,000 tokens of prompt: $4 / $18. A very long campaign that sends its whole history on every turn can cross that line, and then every turn costs double.
It also thinks at the “high” level by default and does not support minimal, so it is the slowest to answer.
It is a preview model from February. Google has announced no shutdown date for it, but preview models are the ones Google replaces: the Gemini 3 Pro preview before it was shut down on 9 March 2026, less than four months after release.
Pick 3.1 Pro if you have run the same scene on Flash and Pro and you can see the difference. If you cannot, stay on Flash.
Gemini 3.5 Flash-Lite — the budget pick
$0.30 / $2.50, a free tier, and minimal thinking by default, so it answers quickly. Google describes it as a model for “high-volume automation”, which is not roleplay. Use it for short scenes, side characters, or a cheap first test of a setup. Gemini 3.1 Flash-Lite is a little cheaper at $0.25 / $1.50, but Google’s table gives it a shutdown date of no sooner than 7 May 2027 and names 3.5 Flash-Lite as its replacement.
Gemini 3.5 Flash and the 2.5 models — move on
Gemini 3.5 Flash is now the worst deal in the list. Google calls it “our legacy Flash model”, and it costs twice as much for input as the three models that replaced it. If your setup still points at gemini-3.5-flash, change it.
Gemini 2.5 Pro and 2.5 Flash are a special case. Google’s change log for 18 September 2026 says:
“we are limiting access to the 2.5 models to users who have actively used them in the past. These models are not deprecated and will continue to be served until further notice through the API. For any new projects, use our latest models: 3.5 Flash-Lite or 3.8 Flash.”
So if you used 2.5 Pro before and like how it writes, you can keep it for now. If you are starting today, it is closed to you, and every guide that recommends it is describing a door that has shut.
What a Long Campaign Costs on Each Model
Price per million tokens is hard to feel. So here is one turn of a mid-campaign session, priced on each model.
The turn: the model receives your system prompt, world notes and chat history — 40,000 tokens in total — and writes a 500-token reply. It is the same turn we priced for Claude and DeepSeek, so the three pages can be compared.
Every turn resends the whole history. Google’s caching guide says implicit caching “is enabled by default for all Gemini 2.5 and newer models”, and that savings are passed on when a request hits the cache. In the “with cache hits” columns, 38,000 tokens are history billed at the cached price and 2,000 are new.
| Model | One turn, no cache hits | One turn, with cache hits | 200 turns, with cache hits |
|---|---|---|---|
| Gemini 3.1 Flash-Lite | $0.011 | $0.002 | $0.44 |
| Gemini 3.5 Flash-Lite | $0.013 | $0.003 | $0.60 |
| Gemini 3.8 / 3.7 / 3.6 Flash (until 31 Dec 2026) | $0.032 | $0.006 | $1.25 |
| The same three, from 1 Jan 2027 | $0.064 | $0.012 | $2.49 |
| Gemini 3.5 Flash | $0.065 | $0.013 | $2.64 |
| Gemini 3.1 Pro | $0.086 | $0.018 | $3.52 |
The numbers are our arithmetic on Google’s price list. Two cautions.
A cache hit is not guaranteed. Google passes the saving on when it happens; it does not promise it on every turn. Without any cache hits, 200 turns on 3.8 Flash cost about $6.40, and on 3.1 Pro about $17.
Thinking is billed as output. Google’s price list says the output price is “including thinking tokens”. The table counts only the 500-token reply. A model thinking at medium or high writes more than that, and you pay for it. This is the practical reason to set thinking to low for roleplay.
Seen next to the other two pages: 200 cached turns cost about $3.50 on Claude Sonnet 5.5, about $1.25 on Gemini 3.8 Flash, and about $0.29 on DeepSeek V4.1 Flash at peak hours. Gemini 3.1 Pro lands almost exactly where Sonnet 5.5 does.
And unlike the other two, Gemini has a free tier on the API. For the whole-category view of subscriptions against per-token pricing, see what an AI RPG actually costs.
The Free API Tier: What You Get and What You Give
“Best free Gemini model for roleplay” is one of the questions people search. On Google’s price list, every Flash and Flash-Lite model shows “Free of charge” in the free-tier column. Gemini 3.1 Pro shows “Not available”.
So the best free Gemini model on the API is the same as the best paid one for most players: Gemini 3.8 Flash.
Two things to know before you rely on it.
Google uses free-tier content. The same price list has a row called “Used to improve our products”. For the free tier it says Yes. For the paid tier it says No. If you would rather your campaign was not used that way, the paid tier is the answer, and at these prices a long session costs about a dollar.
The limits are not published in one place. Google’s rate-limit page says limits depend on your usage tier and “can be viewed in Google AI Studio”. We are not going to quote a number of requests per day, because Google does not publish one in its documentation and the figure changes. Open AI Studio and read the limit for your own account. If a frontend starts returning errors partway through an evening, a free-tier limit is the first thing to check.
Settings That Break Old Gemini Presets
If Gemini has stopped working in a frontend that worked with it before, this is the section to read. Three things changed in 2026.
Temperature, top_p and top_k are deprecated
Google’s change log for 21 July 2026 says: “The sampling parameters temperature, top_p and top_k are now deprecated.” Its migration guide for Gemini 3.8 Flash is direct about what to do:
“Strip temperature, top_p, and top_k from generation configs.”
Even before that, Google’s Gemini 3 guide said it “strongly” recommends keeping temperature at the default of 1.0, and warned that setting it lower “may lead to unexpected behavior, such as looping or degraded performance”.
Most roleplay presets were built around a custom temperature. On current Gemini models, that advice is out of date. Google’s pages do not say what happens to a request that still sends these values, so we will not guess — but its instruction is to remove them. Reset them to default or switch them off, and guide the style with the prompt. Our SillyTavern presets guide explains what each field is.
Prefilled replies
The same migration guide lists “Remove prefilled model turns”. A prefill is a trick many roleplay presets use: the preset writes the first words of the model’s reply for it, to push the tone. If your preset does this and Gemini 3.8 Flash misbehaves, turn the prefill off first.
Thinking levels differ by model
The old thinking_budget number is replaced by thinking_level, and the allowed levels are not the same on every model:
| Model | Default | Levels it accepts |
|---|---|---|
| Gemini 3.8 Flash | Medium | low, medium, high |
| Gemini 3.7 Flash | Medium | low, medium, high |
| Gemini 3.6 Flash | Medium | minimal, low, medium, high |
| Gemini 3.5 Flash-Lite | Minimal | minimal, low, medium, high |
| Gemini 3.1 Pro | High | low, medium, high |
A preset that asks 3.8 Flash for minimal thinking gets an error. Google also notes that even minimal “does not guarantee that thinking is off”.
Using Gemini in SillyTavern, Janitor AI and Other Frontends
A frontend does not change the model. It only changes how you reach it. So the answer is the same as above: 3.8 Flash with thinking on low, or 3.6 Flash for the fastest replies.
You reach Gemini from a frontend in one of two ways: a key from Google AI Studio, or a router such as OpenRouter. On OpenRouter the current models are listed as google/gemini-3.8-flash, google/gemini-3.6-flash and google/gemini-3.1-pro-preview, at the same per-token prices as Google’s. Our OpenRouter roleplay guide covers the setup, and the Janitor AI proxy guide covers where the fields are on Janitor.
One difference from the Gemini app is worth knowing: through the API you get the full 1M-token window on every model, whatever you pay. The 32k limit belongs to the free plan of the app, not to the models. Our context budget calculator shows how far a window goes for a real setup.
Does Gemini Refuse to Roleplay?
No. Gemini plays characters and runs text adventures, and Google’s own safety guide uses “video game dialogue” as its example of a use that may need looser settings.
Here is how the content rules work on the API, in Google’s words:
- There are four adjustable filters: harassment, hate speech, sexually explicit, and dangerous content.
- On current models these are off by default: “the default block threshold is Off for Gemini 2.5 and 3 models”.
- Separately, there are “built-in protections against core harms, such as content that endangers child safety”. These “are always blocked and cannot be adjusted”.
- Google adds: “Applications that use less restrictive safety settings may be subject to review.”
So a blocked reply on a current model is usually not one of the four sliders, because those start switched off. It is either the fixed protections or the model declining by itself, and no setting changes those. Moving from Flash to Pro does not change what Gemini will write either.
If your campaign needs looser content rules, that is a question of which model family to use, not which Gemini. Our guide to uncensored AI roleplay covers those options, and ChatGPT vs Claude vs Gemini for roleplay compares the three big ones.
Our Recommendation
- In the Gemini app, for a long campaign: Google AI Pro, for the 1M-token window. Use Flash day to day and Pro for the scenes that matter.
- In the Gemini app, to try a game: the free plan with a Gem — knowing the 32k window is about 48 pages.
- Through the API, most players: Gemini 3.8 Flash, thinking on low. Newest, free tier, 1M window, and $0.75 / $3.75 until the end of the year.
- Fast, chatty scenes: Gemini 3.6 Flash with minimal thinking.
- Players who prefer its prose: Gemini 3.1 Pro — paid only, about three times the price, and a preview.
- Tight budget: Gemini 3.5 Flash-Lite.
- Leave behind: Gemini 3.5 Flash (costs more than its replacements) and, if you are new, the 2.5 models (closed to new users).
Whatever you pick, the model is only half of the result. A campaign lives or dies on its setup: saved instructions, a world file, and rules for memory. The Gemini RPG guide shows how to build that in a Gem.
Frequently Asked Questions
What is the best Gemini model for roleplay?
It depends on where you play. Through the API, in SillyTavern, Janitor AI or a similar frontend, Gemini 3.8 Flash is the default: it is Google’s newest model (released 2 September 2026), it has a 1M-token context window, and it has a free tier. In the Gemini app you do not get the same list. The free plan gives you 3.6 Flash and varying access to 3.1 Pro, and your plan decides how much of the story the model can see.
Gemini Flash or Pro for roleplay?
Flash for most players. The three newest Flash models cost $0.75 input and $3.75 output per million tokens until 31 December 2026, and all have a free API tier. Gemini 3.1 Pro costs $2 and $12, has no free API tier, is still labelled a preview, and was released in February 2026, so it is older than every Flash model beside it. Choose Pro if you have tried both and prefer its writing in long scenes.
Which Gemini model can I use for free?
On the API, every current Flash and Flash-Lite model has a free tier, including Gemini 3.8 Flash. Gemini 3.1 Pro does not. Google uses free-tier content to improve its products; on the paid tier it does not. In the Gemini app, the free plan includes 3.6 Flash, varying access to 3.1 Pro, and Gems.
How much does Gemini remember in a roleplay?
In the Gemini app it depends on your plan, not the model. Google’s help page lists a context window of 32k tokens without a plan, 128k tokens on Google AI Plus, and 1 million tokens on Google AI Pro and Ultra. Through the API, every current Gemini text model accepts about 1 million tokens of input.
What temperature should I use for Gemini roleplay?
None. Google deprecated the temperature, top_p and top_k settings for its newest models in July 2026, and its guide for Gemini 3.8 Flash says to remove them from your settings. For older Gemini 3 models, Google strongly recommends leaving temperature at the default of 1.0 and warns that lower values can cause looping. Shape the style with your prompt instead.
Can I still use Gemini 2.5 Pro for roleplay?
Only if you already were. On 18 September 2026 Google limited the Gemini 2.5 models to people who have actively used them in the past. Google says they are not deprecated, but new projects are told to use 3.5 Flash-Lite or 3.8 Flash instead. If a guide tells you to start with Gemini 2.5 Pro, it was written before that change.
Does Gemini allow roleplay?
Yes. Gemini plays characters and runs text adventures, and the Gemini app has Gems for saving a game master’s instructions. Through the API, Google’s four adjustable safety filters are off by default on Gemini 2.5 and 3 models. Protections against what Google calls core harms are always on and cannot be changed, and the model can still decline a request on its own.
Which Gemini model is best for Janitor AI or SillyTavern?
Gemini 3.8 Flash with thinking set to low, or Gemini 3.6 Flash if you want the fastest replies, because 3.6 Flash supports the minimal thinking level and 3.8 Flash does not. On OpenRouter they are listed as google/gemini-3.8-flash and google/gemini-3.6-flash at the same prices as Google’s. If an old preset fails, remove the temperature, top_p and top_k values and any prefilled model reply.
Arcanum is independent and not affiliated with Google. Gemini is a trademark of Google LLC, used here only to identify its models.