Chutes AI for Roleplay (2026): Is It Still Free, What It Costs, and How to Set It Up
Chutes is a paid service that runs open-weight AI models, such as DeepSeek, GLM, Kimi and Qwen, and lets apps like Janitor AI, Chub and SillyTavern use them through an API key. Many roleplayers know it as “the free DeepSeek proxy”. That description is out of date. On 11 October 2026 the Chutes pricing page says, in its own words: “We do not offer a free tier at this time.”
This guide explains what Chutes is now, when each free option ended, what a roleplay session costs, how the monthly plans really work, how to connect it, and what the error messages mean.
Everything here comes from Chutes’s own pricing page, its announcements, its documentation and its public source code, all read on 11 October 2026. We did not open a paid account or send requests, so this page makes no claim about speed or writing quality. Where we could not confirm something, we say so.
New to API keys and proxies? Start with what a Janitor AI proxy is. Comparing providers? See OpenRouter vs NanoGPT vs Chutes.
The Short Answer
- Chutes is not free. There is no free tier and no free trial on the pricing page.
- It is cheap for the lightest models. Chutes’s own calculator puts a 30-minute roleplay session at about 5 cents on DeepSeek V4 Flash.
- It hosts 14 text models, all open-weight. No Claude, no GPT, no Gemini.
- There are two ways to pay: top up a balance and pay per token, or buy a monthly plan with a daily request quota.
- The monthly plans have a usage cap that the pricing page does not explain. We explain it below.
- Most “Chutes is broken” moments are one of three things: no balance (error 402), a limit (error 429), or all machines busy (“maximum capacity”).
What Chutes Is
Chutes describes itself as “a serverless AI inference platform”. In plain words: it rents out graphics cards that run AI models, and you pay for what you use.
Two details make it different from most providers:
- The hardware is not in one company’s data centre. Chutes runs on Bittensor subnet 64, which it calls “a decentralized GPU network”. Independent operators, called miners, supply the machines.
- Every hosted model now runs in a protected environment. Chutes calls this TEE, for Trusted Execution Environment. It is the reason every model name ends in
-TEE. Chutes describes the result as “Hardware-protected from everyone—even us”. That is Chutes’s statement. We did not test it.
For a roleplayer the practical part is simple. Chutes gives you an API key and an address. You paste both into your chat app. The app sends your chat to Chutes, and Chutes sends back the reply.
Is Chutes AI Still Free?
No. The pricing page’s FAQ says:
“We do not offer a free tier at this time. Top up your account in the app to get started. Pay-as-you-go has no minimum commitment.”
The free access ended in steps. Chutes’s own posts give the dates.
| Date | What happened | Source |
|---|---|---|
| 24 December 2024 | Chutes launches | Chutes blog |
| 22 January 2025 | DeepSeek R1 becomes available on Chutes | Chutes blog |
| 31 July 2025 | Chutes announces “new pricing tiers, coming next Monday” | Chutes blog |
| 4 August 2025 | The paid plans start | That Monday |
| 27 February 2026 | Chutes announces the end of the Early Access perk: 200 free requests a day. Access to TEE models is removed for those users the same day | Community Announcement |
| 15 March 2026 | The Early Access plan is “fully retired” | Community Announcement |
| 20 March 2026 | Chutes reports that its sponsored free tier on OpenRouter has ended | Business update |
| 15 June 2026 | Chutes reports that every hosted model has moved to TEE and that many models were removed | Chutes blog |
| 11 October 2026 | 14 text models, no free tier | Pricing page and catalog |
Why it ended, in Chutes’s own numbers
Chutes was open about the reason. In its March 2026 update it listed what it had stopped serving:
- About 20 billion tokens a day “from a sponsored free tier via OpenRouter”. Chutes wrote that this program “ended on their timeline” and “was always a limited engagement”.
- About 10 billion tokens a day “from a 200 requests/day free quota offered to ~11,000 users. Average cost: ~$6/user/month. Revenue: zero.”
- “Billions more from subscription abuse and misuse”, including a few companies sending heavy traffic through the $20 plan.
The result, in the same post: “We processed 45% fewer tokens over the last six weeks.”
So the free DeepSeek that older guides describe was real. It was paid for by Chutes and its partners, and they stopped paying for it. If a guide tells you to use a model name ending in :free with a Chutes key, the guide is older than these changes.
If free is what you need, our best free model for roleplay guide lists what is still free today and what the limits are.
The Models and What a Session Costs
Chutes publishes its catalog at a public address, and the pricing page shows the same list. On 11 October 2026 there were 14 text models.
The pricing page includes a calculator with a preset called “30-min roleplay”: about 25 turns of chat with a growing context, 80,000 input tokens and 12,000 output tokens in total. The last column below is Chutes’s own estimate for that session.
| Model | Input, per 1M tokens | Output, per 1M tokens | Context | 30-minute roleplay |
|---|---|---|---|---|
| Mistral Nemo Instruct 2407 | $0.0245 | $0.0978 | 131K | $0.0031 |
| Nemotron 3 Nano Omni 30B | $0.0245 | $0.0978 | 131K | $0.0031 |
| Qwen3 32B | $0.104 | $0.416 | 41K | $0.0133 |
| Gemma 4 31B Turbo | $0.12 | $0.37 | 131K | $0.014 |
| Qwen3 235B Thinking | $0.2989 | $1.1957 | 262K | $0.0383 |
| Qwen 3.8 27B | $0.24 | $2.20 | 262K | $0.0456 |
| Qwen 3.6 27B | $0.30 | $2.00 | 262K | $0.048 |
| DeepSeek V4 Flash | $0.44 | $1.32 | 1M | $0.051 |
| Qwen 3.5 397B | $0.45 | $3.00 | 262K | $0.072 |
| Kimi K2.6 | $0.50 | $2.85 | 262K | $0.0742 |
| DeepSeek V3.2 | $1.00 | $1.00 | 131K | $0.092 |
| GLM-5.1 | $0.98 | $3.08 | 203K | $0.1154 |
| GLM-5.2 | $1.25 | $3.95 | 1M | $0.1474 |
| Kimi K3 | $3.00 | $15.00 | 1M | $0.42 |
Three things to read from this table.
The spread is wide. The same half hour costs a third of a cent on Mistral Nemo and 42 cents on Kimi K3. That is more than a hundred times more. Choosing the model is choosing the bill.
Input is most of the cost. In Chutes’s example, input is 80,000 tokens and output is 12,000. Each turn sends the character card, the lorebook entries and the chat so far. A long chat with a large context costs more per turn than a new one, on every provider. Our token budget guide explains how to keep that number down.
The list is short on purpose. Chutes wrote in June 2026 that it “had to trim down and consolidate many of the models” when it moved everything to TEE. If a model you used last year is missing, that is why. Chutes itself says to read the live list before trusting any model name, because it changes.
Which model for roleplay?
We did not test these models on Chutes, so we will not rank them here. Our model guides do that work from other evidence: best DeepSeek model for roleplay, best model for Janitor AI and best model for SillyTavern.
One fact from the catalog is worth knowing. Chutes publishes the number format each model runs in. This is called quantization, and a smaller format uses less memory. Most models in the list run in 8-bit (fp8). Kimi K2.6 is listed as int4, GLM-5.2 and Gemma 4 31B Turbo as fp4, Kimi K3 as mxfp4, and Qwen3 235B Thinking as bf16. We did not measure whether this changes the writing. We report it because players ask, and because few providers publish it at all.
Two Ways to Pay
Pay as you go
You add money to a balance, and each request takes its token cost from it. The pricing page says there is “no subscription, no minimum, no markup”. For payment, Chutes says it accepts “25+ payment methods via Stripe including crypto, cards, and bank transfers”, and also TAO, the Bittensor token.
For most roleplayers this is the simpler choice. A player who does one 30-minute session a day on DeepSeek V4 Flash spends about $1.50 a month at the prices above.
Monthly plans
The pricing page shows two plans. Chutes’s machine-readable index, which it publishes for AI tools, lists three.
| Plan | Price | Daily request quota | Discount after the quota | Shown on the pricing page |
|---|---|---|---|---|
| Base | $3 a month | 300 | 3% (from the source code) | No |
| Plus | $10 a month | 2,000 | 6% | Yes |
| Pro | $20 a month | 5,000 | 10% | Yes |
We could not see the signed-in billing page, so we cannot confirm whether Base can still be bought by a new account. Check the plan list in your own account before you count on it.
Two rules decide what a plan is worth.
Every call is a request. The pricing page says: “A request is any API call to our service.” A swipe is a request. A regenerate is a request. If your app writes a chat summary for you, that is a request too.
Frontier models may not be in the cheapest plan. In February 2026 Chutes removed “several high-end frontier models” from Base, and named GLM-5, Kimi K2.5, Qwen 3.5 and MiniMax M2.5. The public source code still treats the $3 plan differently for what it calls premium models. Which models count as premium today is not published on the pricing page.
The Usage Cap Nobody Reads
This is the part of Chutes that older guides miss, and it changes what a plan is worth.
Until February 2026 a plan was a simple deal: a fixed number of requests a day, on any model, of any size. Chutes ended that deal because of how it was used. Its announcement showed the heaviest users of each plan. One Pro user had received $6,488 of usage in a month for $20.
The new rule, from the announcement:
“All subscriptions will include a maximum usage allowance equal to 5× the equivalent Pay-As-You-Go value, calculated based on the per-million token pricing of the models used.”
And:
“Limits may apply within shorter rolling time windows (such as per day or per several hours).”
Chutes’s public source code shows the exact numbers that were in use on 3 September 2026. The monthly cap is 5 times the plan price. There is also a cap for each 4-hour window: the plan price, divided by the 180 four-hour windows in a month, times 75.
We did the arithmetic.
| Plan | Monthly cap, in pay-as-you-go value | Cap for each 4-hour window |
|---|---|---|
| Base, $3 | $15 | $1.25 |
| Plus, $10 | $50 | $4.17 |
| Pro, $20 | $100 | $8.33 |
Now put Chutes’s own 30-minute roleplay session into those caps.
| Model | Cost of one session | Sessions inside Plus’s $50 month | Sessions inside one 4-hour window on Plus |
|---|---|---|---|
| DeepSeek V4 Flash | $0.051 | about 980 | about 81 |
| Kimi K2.6 | $0.0742 | about 670 | about 56 |
| GLM-5.2 | $0.1474 | about 340 | about 28 |
| Kimi K3 | $0.42 | about 119 | about 9 |
What this means for a player:
- On light models the cap is far away. Nobody plays 980 half-hour sessions in a month. On DeepSeek V4 Flash, a Plus plan is in practice limited by the 2,000 daily requests, not by the cap.
- On the heaviest model the cap is close. On Kimi K3, Plus covers about four sessions a day, averaged over a month. A long evening with a large context can reach the 4-hour cap.
- A plan is worth at most five times its price. Before February 2026 it could be worth hundreds of times its price. That older deal is the one many guides still describe.
When you pass a cap, Chutes does not stop you if you have a balance. The announcement says: “Once the threshold has been reached, usage will transition to standard Pay-As-You-Go pricing.” With no balance, you get error 402 with the text “Subscription usage cap exceeded. Please add balance to continue.”
One caution. Chutes wrote that “the specific timing and multipliers of these safeguards may evolve over time”. The numbers above are from the public code on one date. Treat them as the shape of the rule, not as a promise.
Plan or pay as you go?
A simple test. Work out what your month would cost on pay-as-you-go, using the session costs above.
- Under the plan price: pay as you go. A plan would cost you more.
- Between the plan price and five times the plan price: the plan saves money, if you stay inside the daily requests.
- Over five times the plan price: you pay the plan price, and then pay-as-you-go rates with a small discount for the rest.
For a casual player on a light model, pay-as-you-go is cheaper. Plans make sense for daily, heavy play on mid-priced models.
Creating an Account: The Fingerprint
Chutes does not sign you in the way most sites do. When you register, it gives you a fingerprint. This is a secret string, and it is the proof that the account is yours. The sign-in page has a field that says “Paste your 32-character fingerprint”.
Chutes’s documentation is direct about it: “Keep Your Fingerprint Safe: Your fingerprint is shown during registration - don’t share it.”
What this means in practice:
- Save it on the day you register. Use a password manager. A screenshot in your photo gallery is not a safe place.
- Do not paste it into a chat app. Chat apps need an API key, which starts with
cpk_. The fingerprint opens the whole account, including the balance. - If you lose it, the sign-in page has a “Forgot fingerprint?” link. We did not go through that process, so we cannot tell you what it asks for.
- “Invalid fingerprint” is the message the source code returns when the string does not match any account. Check for a missing character or an extra space first.
How to Connect Chutes
You need three values, whatever app you use.
| Value | What to enter |
|---|---|
| Address (base URL) | https://llm.chutes.ai/v1 |
| Full chat address, for apps that ask for it | https://llm.chutes.ai/v1/chat/completions |
| API key | Your key from the Chutes dashboard. It starts with cpk_ |
| Model | The exact ID from the catalog, for example deepseek-ai/DeepSeek-V4-Flash-0731-TEE |
Model IDs are case-sensitive strings with the -TEE ending. Copy them. Do not type them from memory.
Janitor AI
Janitor AI calls any outside model a proxy. The steps:
- In Janitor AI, open API Settings.
- Choose Proxy.
- Add a new proxy configuration.
- Enter a name of your choice, the model ID, the full chat address above, and your
cpk_key. - Save, and make sure this configuration is the active one.
- Send a test message.
Our Janitor AI proxy guide covers the proxy screen in detail, and the Janitor AI generation settings guide covers temperature and token limits. Remember that the character must allow proxy use. That is the creator’s setting, not yours.
SillyTavern
SillyTavern 1.19.0 has Chutes built in. You do not need the custom option.
- Open API Connections.
- Set the API to Chat Completion.
- Choose Chutes as the source.
- Paste your
cpk_key and connect. - Pick a model. SillyTavern loads the list from Chutes’s live catalog.
SillyTavern’s code sends the request to the same address as above. If you run an older SillyTavern without the Chutes source, update first. Our SillyTavern presets guide explains which settings apply to a Chat Completion connection.
Chub and other apps
Any app with an “OpenAI-compatible” option can use the three values in the table. Our Chub guide explains where Chub puts its API settings.
Which settings reach the model
For 12 of the 14 models, the catalog lists the sampler settings Chutes accepts: temperature, top P, top K, repetition penalty, frequency penalty, presence penalty, stop strings and seed. A slider outside that list may be ignored. For reasoning models, Chutes has its own guide on how thinking is returned. If your chat shows the model’s reasoning as text, look for a reasoning or thinking switch in your app. Our Janitor AI generation settings guide explains how thinking mode changes the other settings.
Chutes Errors and What They Mean
Chutes’s source code is public, so the messages below are the exact strings, with the status code each one comes with. We read the code as it was on 3 September 2026.
| What you see | Code | What it means | What to do |
|---|---|---|---|
| ”Infrastructure is at maximum capacity, try again later” | Sent inside the reply stream | Every machine running that model is busy | Wait and resend, or change model |
| ”No instances available (yet)“ | 503 | No machine is running that model right now | Wait, or change model |
| ”Rate limit exceeded. Try again later.” | 429 | Too many requests in a short time | Slow down. Check that your key is sent correctly |
| ”Quota exceeded and account balance is $0…“ | 402 | Your plan’s daily requests are used up and you have no balance | Add balance, or wait for the next day |
| ”Subscription usage cap exceeded. Please add balance to continue.” | 402 | You reached the plan’s usage cap and have no balance | Add balance, or wait for the window to pass |
| ”TEE models require an active subscription or positive balance.” | 402 | An older account type with no balance is calling a model it can no longer use | Add balance or buy a plan |
| ”Invalid fingerprint.” | 401 | The fingerprint does not match an account | Check the string. Use “Forgot fingerprint?” |
A few notes on the ones players meet most.
”Infrastructure is at maximum capacity”
This is not about you. Chutes’s machines for that model are all in use. The code logs it as “all miners are at max capacity”. It is most likely at busy hours and on the most popular models. Chutes itself wrote in February 2026 that some frontier models were “frequently overloaded” and that it was “limited due to a global shortage” of graphics cards.
Because the message arrives inside the reply stream, some apps do not show it as an error. They show an empty reply, or a “no response” notice. If Chutes replies with nothing at a busy hour, this is a likely cause.
Error 429 when you have not sent much
Chutes’s documentation gives one cause that is easy to miss. The key must be sent as a Bearer token. If the app sends it another way, the request is “treated as unauthenticated” and can “hit the anonymous rate-limit path”. In simple terms: Chutes does not recognise you, and treats you as a stranger with a very small limit. If 429 appears on your first message, check the key field for a missing character, and check that you chose the right connection type.
Error 402
402 means “payment required”. On Chutes it has three causes, and adding balance fixes all of them. The most common one today is a player following an old guide: an account with no plan and no balance, calling a model that used to be free.
Is Chutes down?
Chutes has a status page at status.chutes.ai. It tracks the website and the API. When we read it on 11 October 2026 it showed all services online. The page does not show whether one model is at capacity, so a green status page and a “maximum capacity” message can both be true at the same moment.
For errors that come from the app and not from Chutes, see Janitor AI not working, SillyTavern not working and Chub AI not working.
Privacy
Three facts, each from Chutes’s own pages.
- Every model in the catalog is marked as confidential compute. Chutes’s own summary is: “Your prompts. Your data. Hardware-protected from everyone—even us.”
- The machines belong to independent operators. Chutes describes them as “anonymous and permissionless”, and says this is why it built the protection: “don’t trust, verify”.
- Chutes also offers an optional end-to-end encrypted path. Chat apps like Janitor AI do not use it by default.
We did not audit any of this. We can say that Chutes publishes the code and the method, and that few providers do. We cannot say more than that. And remember the other half of the path: your chat app also sees your messages. Our data custody audit explains what the chat platforms themselves keep.
Should You Use Chutes?
Chutes fits you if:
- You want open-weight models such as DeepSeek, GLM, Kimi or Qwen, and you are happy with a short list.
- You like paying a few cents per session with no monthly fee.
- The privacy design matters to you.
Look elsewhere if:
- You want free. Chutes is not free. See best free model for roleplay.
- You want Claude, GPT or Gemini. Chutes does not host them. See OpenRouter for roleplay.
- You want hundreds of models and community fine-tunes behind one key. See NanoGPT for roleplay.
For a side-by-side view of the three providers roleplayers ask about most, read OpenRouter vs NanoGPT vs Chutes.
Frequently Asked Questions
Is Chutes AI free?
No. On 11 October 2026 the Chutes pricing page says: We do not offer a free tier at this time. You pay for the tokens you use, or you buy a monthly plan. The free DeepSeek access that made Chutes popular with roleplayers ended in steps between August 2025 and March 2026.
When did Chutes stop being free?
In three steps. Paid plans arrived on 4 August 2025. On 27 February 2026 Chutes announced the end of its Early Access perk of 200 free requests a day, which was fully retired on 15 March 2026. In March 2026 Chutes also reported that the sponsored free tier it served through OpenRouter had ended.
How much does Chutes cost for roleplay?
Chutes’s own calculator prices a 30-minute roleplay session, about 25 turns with 80,000 input tokens and 12,000 output tokens, at about 5 cents on DeepSeek V4 Flash, 7 cents on Kimi K2.6, 15 cents on GLM-5.2 and 42 cents on Kimi K3. These are pay-as-you-go prices read on 11 October 2026.
What is the Chutes proxy URL for Janitor AI?
The chat completions address is https://llm.chutes.ai/v1/chat/completions. In Janitor AI’s proxy settings you enter that address, your Chutes API key, which starts with cpk_, and the exact model ID from Chutes’s model list, for example deepseek-ai/DeepSeek-V4-Flash-0731-TEE.
What does Infrastructure is at maximum capacity mean on Chutes?
It means every machine running that model is busy. Chutes’s public source code sends this message when no machine can take the request. Your key and your balance are not the problem. Wait and send again, or switch to a model with more free capacity.
Why does Chutes show error 402?
Error 402 means payment is required. In Chutes’s source code it appears when your daily plan quota is used up and your balance is zero, when your plan’s usage cap is reached and your balance is zero, or when you have no plan and no balance. Adding balance clears all three.
Does Chutes have Claude, GPT or Gemini?
No. Chutes hosts open-weight models only. On 11 October 2026 its public catalog listed 14 text models from DeepSeek, Z.ai, Moonshot AI, Qwen, Google’s Gemma family, NVIDIA and Mistral. For Claude, GPT or Gemini you need a different provider.
What is a Chutes fingerprint?
It is the secret that proves an account is yours. Chutes shows it once when you register, and the sign-in page accepts a 32-character fingerprint. Chutes’s own documentation says to keep it safe and not to share it. Save it in a password manager on the day you register.
Prices, the model list, the roleplay estimate, the plan prices, the payment methods and the “no free tier” statement are from the Chutes pricing page, its public model catalog and its published index for AI tools, read on 11 October 2026. The dates and figures for the end of free access are from Chutes’s own posts of 31 July 2025, 27 February 2026, 20 March 2026 and 15 June 2026. The usage cap numbers and the error messages are from Chutes’s public API source code as it stood on 3 September 2026. The SillyTavern steps are from SillyTavern 1.19.0’s code. We did not open a paid Chutes account or send requests, and nothing on this page is a first-hand test of speed or quality. Arcanum has no commercial relationship with Chutes.
Chutes, Bittensor, Janitor AI, SillyTavern, Chub, OpenRouter, NanoGPT, DeepSeek, Kimi, GLM and Qwen are the property of their respective owners. Arcanum is not affiliated with any of them.