OpenRouter vs NanoGPT vs Chutes (2026): Which API Provider for Roleplay?
OpenRouter, NanoGPT and Chutes are three services that sell access to AI models through an API key, so that apps like Janitor AI, SillyTavern and Chub can use a better model than the built-in one. Players compare them all the time, and most comparisons are about price.
Price is harder to compare than it looks, because the three charge in different units. OpenRouter charges a fee when you add money. NanoGPT’s subscription counts input tokens per week. Chutes’s plans count requests per day, under a cap measured in dollars.
This page puts all three on the same example. Everything comes from each provider’s own pricing page, documentation and public model list, read on 11 October 2026. We did not open paid accounts or send requests. There is no speed table and no quality ranking here, because we did not test either.
Each provider has its own guide: OpenRouter for roleplay, NanoGPT for roleplay and Chutes AI for roleplay.
The Short Answer
- Choose OpenRouter if you want free models, or if you want the provider that most apps and guides assume.
- Choose NanoGPT if you want the lowest list prices on open-weight models, a very small first deposit, or a fixed monthly bill.
- Choose Chutes if you want open-weight models that all run in a protected environment, and a short list is enough.
- For Claude, GPT or Gemini, use OpenRouter or NanoGPT. Chutes does not have them.
- You can use more than one. They are separate keys, and your chat app can save several connections.
Side by Side
| OpenRouter | NanoGPT | Chutes | |
|---|---|---|---|
| What it is | Aggregator: one key for many providers | Aggregator: one key for many providers | Its own network of machines |
| Models in the list we read | 458 | 643 | 14 |
| Claude, GPT, Gemini | Yes | Yes | No |
| Community roleplay fine-tunes | Some | Many | No |
| Free models for API use | Yes, 15 entries | No (one free model, website only) | No |
| Fee when you add money | 5.5%, $0.80 minimum (card); 5% (crypto) | None stated | None stated |
| Markup on model prices | None stated | None stated; 5% if you pin a provider | Sets its own prices |
| Smallest deposit | Not checked | $0.10 crypto, $1 card | ”No minimum” stated |
| Subscription | None for individuals | $12 a month | $10 and $20 a month shown |
| What the subscription counts | — | Input tokens per week | Requests per day, under a dollar cap |
| Account needed | Yes | No, by its own statement | Yes, with a fingerprint |
| Built into SillyTavern 1.19.0 | Yes | Yes | Yes |
“None stated” means the provider says it does not charge this. “Not checked” means we did not find the figure on the pages we read.
How Each One Charges
OpenRouter: a fee on the way in
OpenRouter says it passes model prices through “without any markup”. It earns its money when you buy credits. Its documentation gives the fee as 5.5% with a $0.80 minimum for card payments, and 5% for crypto.
The minimum matters for small purchases. A fee of 5.5% is less than $0.80 on anything under about $14.50, so the minimum applies.
| Card purchase | Fee | Fee as a share |
|---|---|---|
| $10 | $0.80 | 8% |
| $20 | $1.10 | 5.5% |
| $50 | $2.75 | 5.5% |
A common purchase is $10, because buying 10 credits raises the daily limit on free models. On that purchase the fee is 8%. It is still only 80 cents, but it is not the 5.5% that most guides quote.
NanoGPT: list price, and an optional weekly allowance
NanoGPT says it charges “API at list prices — no markup” and “no percentage fees on deposits”. The smallest deposit is $0.10 with crypto or $1 with a card.
Its optional subscription is $12 a month for 60 million input tokens a week on a list of included models. Our NanoGPT guide explains what that list contains.
Chutes: its own prices, and plans with a cap
Chutes is not an aggregator. It runs the models on its own network and sets its own prices. The pricing page says: “No subscription, no minimum, no markup.”
Its optional plans give a daily number of requests. Since February 2026 each plan also has a usage cap of five times the plan price. Our Chutes guide explains the cap in detail.
What One Session Costs
To compare, we use one example on all three: 30 minutes of roleplay, about 25 turns, with 80,000 input tokens and 12,000 output tokens in total. The example is from Chutes’s pricing calculator. The arithmetic is ours, from each provider’s listed prices on 11 October 2026.
| Model | OpenRouter | NanoGPT | Chutes |
|---|---|---|---|
| DeepSeek V4 Flash (0731) | $0.009 | $0.006 | $0.051 |
| Qwen 3.8 27B | $0.031 | $0.020 | $0.046 |
| DeepSeek V3.2 | $0.030 | $0.027 | $0.092 |
| GLM 5.2 | $0.089 | $0.049 | $0.147 |
| Kimi K2.6 | $0.093 | $0.071 | $0.074 |
| Kimi K3 | $0.184 | $0.28 | $0.42 |
| Claude Sonnet 5.5 | $0.28 | $0.28 | Not offered |
Read this table with three cautions.
OpenRouter’s number is a listed price, not a fixed one. OpenRouter sends each request to one of several providers, and they do not all charge the same. Its model list shows one price per model. What you pay depends on which provider serves you. The table also leaves out the fee on credits, which adds 5.5% to 8% for most roleplayers.
Chutes’s number buys something the others do not. Every Chutes model runs in a protected environment, called TEE. The OpenRouter and NanoGPT prices above are for standard runs. NanoGPT also sells protected versions of some of these models, with names that start with TEE/. Against those, Chutes looks different:
| Model, protected version | NanoGPT TEE/ | Chutes |
|---|---|---|
| Qwen 3.8 27B | $0.068 | $0.046 |
| DeepSeek V3.2 | $0.052 | $0.092 |
| GLM 5.2 | $0.165 | $0.147 |
| Kimi K2.6 | $0.183 | $0.074 |
| Kimi K3 | $0.42 | $0.42 |
So Chutes is the expensive one if you only want the model, and often the cheaper one if you want the protected run.
Prices move. All of these were read on one day. Check the live page before you decide.
What the table does show clearly:
- No provider is cheapest on everything. NanoGPT is lowest on five of the six open-weight rows. OpenRouter is lowest on Kimi K3.
- On Claude there is nothing to choose on price. Both aggregators list the same price. The difference is OpenRouter’s fee on credits.
- Every number here is small. The largest is 42 cents for half an hour. The choice of model changes your bill far more than the choice of provider does.
Most of each cost is input, because every turn sends the whole context again. Our token budget guide explains how to bring it down.
What Is Still Free
| Free for use in a chat app? | The limit | |
|---|---|---|
| OpenRouter | Yes. 15 entries ending in :free on 11 October 2026 | 20 requests a minute. 50 a day, or 1,000 a day once you have bought at least 10 credits |
| NanoGPT | No. One free model, on the website only | — |
| Chutes | No. “We do not offer a free tier at this time” | — |
OpenRouter is the only free route of the three. Not all 15 entries are useful for stories. Several are coding or safety models. Our best free model for roleplay guide goes through the list and names the ones players use.
Two details from OpenRouter’s documentation:
- The limit counts requests. Every swipe and every regenerate is one request.
- A negative balance blocks free models too. OpenRouter says that with a negative credit balance you may see error 402 “including for free models”.
Chutes used to be the free option. Its free access ended in steps between August 2025 and March 2026, and the Chutes guide has the dates. If an older guide sends you to Chutes for free DeepSeek, that guide is out of date.
The Two Subscriptions
OpenRouter has no subscription for individual players. NanoGPT and Chutes both have one, and they count different things.
| NanoGPT Pro | Chutes Plus | Chutes Pro | |
|---|---|---|---|
| Price | $12 a month | $10 a month | $20 a month |
| What it counts | 60 million input tokens a week | 2,000 requests a day | 5,000 requests a day |
| Extra cap | None stated on the pricing page | $50 of usage a month | $100 of usage a month |
| Models it covers | A list of 311 entries | Chutes’s catalog | Chutes’s catalog |
| After the limit | Balance, if the key allows it | Balance, with a 6% discount | Balance, with a 10% discount |
| Also includes | 100 images a day | — | — |
Chutes’s usage caps are five times the plan price. That rule is from its announcement of 27 February 2026, and the numbers match its public source code as of 3 September 2026. Chutes says the numbers “may evolve over time”.
Because one plan counts tokens and the other counts requests and dollars, the fair way to compare is to pick a model and a context size and count turns. Here is DeepSeek V4 Flash, which both plans cover, with a reply of about 480 tokens per turn.
| Input sent per turn | NanoGPT Pro, turns a day | Chutes Plus, turns a day |
|---|---|---|
| 3,200 tokens | about 2,680 | about 815 |
| 20,000 tokens | about 430 | about 175 |
| 50,000 tokens | about 170 | about 75 |
How we got these: NanoGPT’s column is 60 million tokens, divided by the input per turn, divided by 7 days. Chutes’s column is the $50 monthly cap, divided by the pay-as-you-go cost of one turn at Chutes’s prices, divided by 30 days.
On this model, NanoGPT’s plan goes further at every size. On a more expensive model the gap grows, because Chutes’s cap is in dollars and NanoGPT’s allowance is in tokens. A token on Kimi K2.6 uses the NanoGPT allowance at the same rate as a token on a cheap model. On Chutes it uses the cap faster.
Two things the table does not show:
- Kimi K3 is in Chutes’s catalog. It is not in NanoGPT’s included list. If that is your model, the comparison changes.
- Neither plan is needed by a casual player. One session a day on DeepSeek V4 Flash costs about 18 cents a month on NanoGPT’s pay-as-you-go and about $1.50 on Chutes’s. Both are far under the plan price.
The Models
OpenRouter listed 458 models in its public list on 11 October 2026. Its pricing page describes “500+ models” from “80+ providers”. It has the commercial models, the large open-weight models, and the free entries.
NanoGPT listed 643. On top of the commercial and open-weight models, it has a large number of community fine-tunes made for fiction and roleplay, and “uncensored” versions of several models. If you want a specific fine-tune by name, this is the first place to look.
Chutes listed 14. All are open-weight: DeepSeek V4 Flash and V3.2, GLM-5.1 and 5.2, Kimi K2.6 and K3, five Qwen models, Gemma 4 31B, Mistral Nemo and a small NVIDIA model. Chutes cut its list in 2026 when it moved everything to the protected environment.
For which model to pick, see best LLM for roleplay, best DeepSeek model for roleplay and best Claude model for roleplay. Those guides work from other evidence. This page does not rank models.
Privacy
With any of the three, two parties see your chat: the chat app and the provider. With an aggregator there is a third: the company that runs the model.
- OpenRouter publishes a table of its providers that shows, for each one, how long prompts are kept and whether the provider may train on them. The policies differ from provider to provider.
- NanoGPT says in its privacy policy: “By default, we do NOT store prompt or conversation content on our servers.” It also says the prompt is passed to the model’s provider, which “may store and process this information according to their own privacy policies”. Its policy lists what individual providers do.
- Chutes is built differently. Every model is marked as confidential compute, and Chutes’s own summary is: “Hardware-protected from everyone—even us.” The machines belong to independent operators that Chutes describes as “anonymous and permissionless”, which is the reason it built the protection.
We did not audit any of these claims. What we can say is where each one puts its promise. OpenRouter and NanoGPT tell you which providers keep what, and let you choose. Chutes says the hardware prevents anyone from reading the content.
For OpenRouter’s free models there is one more thing to know: free often has a privacy cost. Our free model guide has a section on it. And for what the chat apps themselves keep, see the data custody audit.
Connecting Each One
All three use the same kind of connection, called OpenAI-compatible. You need an address, a key and a model ID.
| Chat address | Key starts with | |
|---|---|---|
| OpenRouter | https://openrouter.ai/api/v1/chat/completions | sk-or- |
| NanoGPT | https://api.nano-gpt.com/api/v1/chat/completions | sk-nano- |
| Chutes | https://llm.chutes.ai/v1/chat/completions | cpk_ |
In Janitor AI, you add each one as a proxy configuration: a name, the model ID, the chat address and the key. Our Janitor AI proxy guide walks through the screen. The character must allow proxy use, which is the creator’s choice.
In SillyTavern 1.19.0, all three are built in. Open API Connections, choose Chat Completion, pick the source by name, paste the key, and choose a model from the list that loads.
Model IDs are different on each provider, even for the same model. DeepSeek V4 Flash is deepseek/deepseek-v4-flash-0731 on OpenRouter and NanoGPT, and deepseek-ai/DeepSeek-V4-Flash-0731-TEE on Chutes. Copy the ID from the provider you are using. An ID from another provider’s guide is the most common reason for a “model not found” error.
Limits and Errors
The same error numbers mean slightly different things on each.
| Code | OpenRouter | NanoGPT | Chutes |
|---|---|---|---|
| 402 | Not enough credits, or a negative balance | Not enough balance | Plan quota or cap reached with no balance |
| 429 | A free-model limit, or the provider behind it is limiting you | 25 requests a second, or your own daily key limit | Too many requests, or the key was not recognised |
| 503 or a capacity message | The provider is unavailable; OpenRouter tries the next one | The model is temporarily unavailable | ”Infrastructure is at maximum capacity” |
Three differences are worth knowing before you choose.
- OpenRouter falls back. Its documentation says that if a provider returns an error, it “will automatically fall back to the next provider”. With one model served by several companies, one of them being down is often invisible to you.
- NanoGPT lets you set your own ceiling. Each key can have a limit of requests per day or dollars per day. For a player who worries about a surprise bill, this is the most useful control of the three.
- Chutes has one network. When all machines for a model are busy, there is no second provider to try. Chutes wrote in February 2026 that some large models were “frequently overloaded”. We did not measure how often this happens now.
For errors that come from the app, see Janitor AI not working, SillyTavern not working and Chub AI not working.
Which One Should You Use?
You want to spend nothing. OpenRouter. It is the only one with free models for chat apps. Expect 50 requests a day.
You want the cheapest good DeepSeek or GLM chat. NanoGPT on pay-as-you-go had the lowest listed prices on those models on the day we read them. OpenRouter is close on some and has the fee on credits.
You want Claude. OpenRouter or NanoGPT. Same listed price. NanoGPT states no fee on deposits.
You play for hours every day and want a fixed bill. NanoGPT’s $12 plan, if your model is in its included list. Check the list first.
You care most about who can read your chat. Chutes, for the design it describes. Read its own pages and decide whether you accept the claim. We did not test it.
You are following a guide. Use the provider the guide uses. Most Janitor AI and SillyTavern guides assume OpenRouter, and matching the guide saves time on model IDs and settings.
You cannot decide. Start with OpenRouter’s free models to learn how a key works. When you want a better model, put one dollar into NanoGPT and try it. The cost of being wrong is very small with any of the three.
We did not research other providers for this page. Featherless and direct keys from DeepSeek, Anthropic or Google are real alternatives, and they are not compared here. Our Janitor AI API key guide covers direct keys.
Frequently Asked Questions
Which is better for roleplay, OpenRouter, NanoGPT or Chutes?
It depends on what you need. OpenRouter is the only one of the three with free models for API use, and it is the provider most apps and guides assume. NanoGPT has the lowest list prices on most of the open-weight models we compared and an optional $12 subscription. Chutes has a short list of open-weight models that all run in a protected environment.
Is OpenRouter, NanoGPT or Chutes free?
Only OpenRouter has free models you can use from Janitor AI or SillyTavern. On 11 October 2026 its model list had 15 free entries, limited to 50 requests a day, or 1,000 a day once you have bought at least 10 credits. NanoGPT’s one free model works on its website only. Chutes says it does not offer a free tier.
Does OpenRouter charge a fee?
Yes, when you buy credits. OpenRouter’s documentation gives 5.5% with a $0.80 minimum for card payments and 5% for crypto. It says it adds no markup to the model prices themselves. On a $10 card purchase the minimum applies, so the fee is $0.80, which is 8%.
Which is cheapest for DeepSeek roleplay?
On list prices read on 11 October 2026, NanoGPT. Our example 30-minute session on DeepSeek V4 Flash came to about $0.006 on NanoGPT, $0.009 on OpenRouter and $0.051 on Chutes. Chutes’s price is for a run in a protected environment, which the other two prices are not.
Which has a subscription, and which one goes further?
NanoGPT has one plan at $12 a month with 60 million input tokens a week. Chutes shows two plans, $10 and $20 a month, with daily request quotas and a usage cap of five times the plan price. OpenRouter has no subscription for individuals. On DeepSeek V4 Flash, our arithmetic gives NanoGPT’s plan more turns a day than Chutes’s $10 plan at every context size we checked.
Which one has Claude, GPT and Gemini?
OpenRouter and NanoGPT both have them, on pay-as-you-go, at the same list price for the models we compared. Chutes does not host them. Chutes hosts open-weight models only.
Do all three work with Janitor AI and SillyTavern?
Yes. Janitor AI connects to any of them through its proxy settings, with a chat address, a key and a model ID. SillyTavern 1.19.0 has all three built in as Chat Completion sources, so you only need to paste a key and pick a model.
Can I use more than one?
Yes. Each is a separate account and a separate key, and chat apps let you save more than one connection. A common setup is OpenRouter for its free models and a paid key from another provider for the model you use most.
OpenRouter’s fees, free-model limits, fallback behaviour and error descriptions are from its documentation, and its model count and prices are from its public model list. NanoGPT’s prices, subscription, deposit minimums, key limits and privacy quotes are from its pricing page, documentation, public model lists and privacy policy. Chutes’s prices, plans and statements are from its pricing page, public model catalog and announcements, and its usage cap numbers are from its public source code as of 3 September 2026. Everything else was read on 11 October 2026. The session costs, the fee table and the turns-per-day table are our own arithmetic from those figures. We did not open paid accounts with any of the three, we did not send requests, and nothing on this page is a first-hand test of speed, reliability or writing quality. Arcanum has no commercial relationship with OpenRouter, NanoGPT or Chutes.
OpenRouter, NanoGPT, Chutes, Janitor AI, SillyTavern, Chub, Claude, GPT, Gemini, DeepSeek, Kimi, GLM and Qwen are the property of their respective owners. Arcanum is not affiliated with any of them.