Published Updated
In one or two sentences, before anything else.
Chutes is a decentralized inference network on Bittensor subnet 64, run by Chutes Global Corp, where anonymous GPU operators serve open-weight models such as DeepSeek, Kimi and GLM through an OpenAI-compatible API. It ended its free tier on March 15, 2026, runs all 14 public models in trusted execution environments, and relies on TAO token subsidies.
What it can do, one sentence each, and whether we actually tested it.
All 14 public LLMs, including DeepSeek V4 Flash, Kimi K3 and K2.6, GLM-5.2 and Qwen3.8 27B, are marked as running on Intel TDX machines with NVIDIA confidential-computing GPUs.
Andrew Miller's October 2026 audit found the attestation core sound, but the serving code and model are not part of the attestation.
Pay-as-you-go with no minimum, or Plus at 10 USD a month for 2,000 requests a day and Pro at 20 USD for 5,000 a day.
Plans are capped at 50 and 100 USD of pay-as-you-go value a month, so the daily count is often not reachable, and Chutes' terms limit them to casual, direct use.
A request can name a saved pool such as default:latency or an inline list such as modelA,modelB:latency, and Chutes picks among them.
Failover is only between Chutes' own models; there is no fallback to another host.
An /e2e/invoke path encrypts prompts with post-quantum ML-KEM-768 so that Chutes' API servers cannot read them.
Andrew Miller's audit found the shipped client does not check the key against attestation and streamed chunks have no integrity check; the default path is plaintext at Chutes' API.
A chutes-ai plugin for Claude Code and Cowork with nine SKILL.md skills (routing, TEE checks, usage and billing, deployment and more) and a local MCP server for models, quota, usage and keys.
Documented on chutes.ai but kept in the Veightor/chutes-agent-toolkit GitHub repository, not Chutes' chutesai organization; most write tools and six of the nine skills are marked beta.
An OAuth sign-in that lets third-party apps bill model use to each user's own Chutes account.
Per Chutes' llms.txt.
Stripe with more than 25 methods, including cards, bank transfers and crypto, plus direct TAO and Bittensor alpha-token deposits.
Chutes' terms make crypto payments final and non-refundable.
The people it helps and the problems it solves, and where it ranks.
Built for
Problems it solves
Graded in
Public tests other people ran: what they gave it and how it did.
Case 1
Input
About 4,000 tool-calling requests replayed against Chutes' Kimi K2 Thinking and K2-0905 endpoints and compared with Moonshot's own API.
What happened
On K2 Thinking, Chutes had the lowest scores of 17 entries: 68.10% trigger similarity, below Moonshot's 73% line, and 83.05% valid schemas (3,037 of 3,657) against 100% on Moonshot's API. On K2-0905 it passed with 96.70%. The test predates Chutes' TEE-only catalog.
Case 2
Input
For all 14 confidential-compute models, verify the TDX quote, nonce and key binding, measurements and NVIDIA GPU evidence, then send an end-to-end encrypted request.
What happened
Attestation and key binding checked out for 10 models. One Qwen3-32B instance ran an unpublished image, three models gave no attestation evidence and GLM-5.2 returned 429. Issues he filed with Chutes on June 18, 2026 were still open with no reply.
Case 3
Input
Weeks of OpenCode coding and a swarm of grammar-checking subagents on Chutes' 3 USD and 10 USD plans, compared with Synthetic.new and Z.ai.
What happened
He found the speed adequate but slower than Synthetic (about three times slower on Kimi K2.5 the first evening), sessions occasionally stalled, and he called it "probably the best value". Informal timings, written before the 2026 plan and catalog changes.
Case 4
Input
A Janitor AI proxy pointed at https://llm.chutes.ai/v1/chat/completions with a Chutes API key and a model ID.
What happened
The help article lists Chutes' outage errors: 500 "exhausted all available targets" when Chutes is down and 503 "No instances available" when a model is overloaded or restarting. Free DeepSeek guides from 2025 no longer apply.
What it does badly, with the evidence, and what we could not settle.
Weak
No free tier, and prices have risen.
Evidence: The free tier ended on March 15, 2026 and the 3 USD Base plan is gone. Robo Rhythms, an affiliate site, reports DeepSeek V3.2 rising from 0.24 / 0.42 USD to 1.00 / 1.00 USD per million tokens without notice; the current list shows 1.00 / 1.00 USD.
Busy models fail instead of falling back.
Evidence: OpenRouter's one-day uptime for Chutes on October 21, 2026 was 86.32% on Qwen3.6 27B, 89.44% on GLM-5.1 and 90.74% on Kimi K3. Chutes said in February 2026 that large models were frequently overloaded, and there is no fallback to another host.
Last place on Moonshot's Kimi K2 Thinking tool-call test.
Evidence: 68.10% trigger similarity and 83.05% schema accuracy in November 2025, the lowest of 17 entries.
The privacy story has documented gaps.
Evidence: Andrew Miller's audit found the default API path is plaintext at Chutes' control plane, the attestation does not cover which model or code handled a prompt, and a cross-user prompt leak from a verified enclave was demonstrated in June 2026.
Prices depend on a shrinking token subsidy.
Evidence: Pine Analytics estimated in March 2026 that SN64 received about 52 million USD a year in TAO emissions against 1.3 to 2.4 million USD of customer revenue. Chutes' emission share fell from about 14.39% to 5.82% by September 22, 2026, and its public revenue series stopped after September 6.
Not for production apps or buyers who need a conventional company.
Evidence: Plans are banned for automated or production use, balances may be forfeited on termination, the terms follow Nevis law with arbitration and a class-action waiver, and capacity errors have no fallback. Apps that need closed models, guaranteed capacity or a company accountable under US or EU law should look elsewhere.
Uncertain
How fast Chutes is.
Why unsure: Chutes does not appear on 16 Artificial Analysis provider pages for models it serves or has served, so no independent speed benchmark has been published.
How much traffic it really serves.
Why unsure: Rayon Labs claims 160 billion tokens a day; Own Your Mind reports OpenRouter-measured Chutes volume of about 7 billion a day by April 2026.
The minimum deposit.
Why unsure: The pricing page publishes no minimum; Hacker News users describe a 5 USD deposit.
Whether tool-call accuracy changed after the move to TEEs.
Why unsure: Moonshot's results date from November 2025, and the newer Kimi Vendor Verifier publishes no vendor results.
The numbers, each with its source and the date we checked it.
| Spec | Value | Note and source |
|---|---|---|
| Operator | Chutes Global Corp, under the laws of Nevis, Saint Kitts and Nevis; built by Rayon Labs as Bittensor subnet 64 | Chutes says it has no CEO and is run by mostly pseudonymous contributors. Source: chutes.aichecked |
| Kimi K2.6 on Chutes, per million tokens | 0.45 USD input, 1.95 USD output | Moonshot AI's own endpoint on OpenRouter: 0.95 / 4.00 USD. Source: chutes.aichecked |
| Platform fee | None ("No subscription, no minimum, no markup") | Source: chutes.aichecked |
| Plans and minimum | No published minimum; Plus 10 USD a month (2,000 requests a day), Pro 20 USD (5,000 a day) | Plans are capped at 50 and 100 USD of pay-as-you-go value a month. Source: chutes.aichecked |
| Payment methods | Stripe (cards, bank transfers, crypto), TAO and Bittensor alpha-token deposits; invoicing for enterprise | Source: chutes.aichecked |
| Free tier | None since March 15, 2026 | Source: chutes.aichecked |
| Models | 14 public LLMs, all open-weight and all in TEEs; team-published image, video and music models | Quantization varies: fp8 for most, fp4, int4 or mxfp4 for some. Source: llm.chutes.aichecked |
| API formats | OpenAI-compatible Chat Completions with streaming, tools, JSON mode, structured outputs and vision; no Anthropic Messages or Responses endpoint documented | Source: chutes.aichecked |
| MCP | Local stdio MCP server in its agent toolkit: models, quota, usage, aliases and keys; write and deploy tools in beta | Source: chutes.aichecked |
| Agent Skills | chutes-ai plugin for Claude Code and Cowork with nine SKILL.md skills | Hosted in the Veightor/chutes-agent-toolkit GitHub repository. Source: chutes.aichecked |
| Coding agents | Guides for Claude (plugin), Cursor, Cline, Aider, Hermes, OpenClaw (beta) and Codex (guide only); built into SillyTavern | Source: chutes.aichecked |
| Privacy and data retention | No logging or storing of prompts and responses for public LLM inference; no training on user data | In standard mode servers may read request data in memory; Chutes' own chat apps store conversations. Source: chutes.aichecked |
| Status page and uptime | status.chutes.ai shows 100% over 30 days for the website and an API heartbeat; OpenRouter one-day uptime 86% to 100% by model | Source: status.chutes.aichecked |
| Refunds | Crypto payments final; Stripe refunds only in limited subscription cases at Chutes' discretion; balance may be forfeited on termination | Source: chutes.aichecked |
Answer first, then why, an example, the limits and what to do next.
Chutes AI no longer has a free tier. Its 200-requests-a-day Early Access tier ended on March 15, 2026, and the pricing page says it does not offer a free tier; the cheapest options are pay-as-you-go with no published minimum or Plus at 10 USD a month.
Why
Chutes said in March 2026 that it cut free and subsidised usage to build a sustainable business.
Chutes undercuts the model makers' own prices on the models they both sell. Kimi K2.6 costs 0.45 / 1.95 USD per million tokens on Chutes against 0.95 / 4.00 USD from Moonshot AI, about half, and GLM-5.1 costs 0.98 / 3.08 USD against Z.AI's 1.40 / 4.40 USD, about 30% less.
Limits
Those prices are subsidised by TAO emissions that are shrinking, and Chutes raised some prices in 2026.
Chutes AI runs every public model in a TEE and bans prompt logging, but its operator is opaque and an independent audit found gaps. Andrew Miller's October 2026 audit found the default API path plaintext at Chutes and the serving code and model unattested.
Limits
The company is registered under Nevis law, its team is mostly pseudonymous, and balances may be forfeited if an account is ended.
A 503 "No instances available" error from Chutes means the model is overloaded or restarting, per the Janitor AI help center. A 500 "exhausted all available targets" means Chutes itself is down; in both cases the only fix is to wait or switch models, because Chutes has no fallback host.
Chutes documents a local MCP server and a Claude Code plugin with nine skills. The MCP server reads models, quota, usage and keys, with write tools in beta; Claude stays the model and calls Chutes through tools, because Chutes documents no Anthropic-compatible endpoint.
Novita AI, NanoGPT and OpenRouter are the usual alternatives for cheap open models. Novita charges more (Kimi K2.6 at 0.80 / 3.40 USD) with zero retention by default; NanoGPT's 12 USD PRO plan covers open models up to 60 million input tokens a week; OpenRouter routes between hosts with fallback.
This page is built from Chutes' own site, docs and terms, OpenRouter's endpoint data, Moonshot AI's K2 Vendor Verifier, Andrew Miller's audit, analyses of its token economics and user reports, all checked on October 21, 2026. Chutes' claims are labelled as its own.
Chutes is an inference network on Bittensor subnet 64, launched in December 2024 and built by Rayon Labs. Anonymous GPU operators, called miners, run the models, and Chutes sells access through an OpenAI-compatible API at https://llm.chutes.ai/v1. Its terms name Chutes Global Corp under the laws of Nevis, Saint Kitts and Nevis. Chutes says it has no CEO and is run by mostly pseudonymous contributors.
The public LLM list had 14 models on October 21, 2026, all open-weight and all marked as running in trusted execution environments (TEEs): DeepSeek V4 Flash 0731 and V3.2, Kimi K3 and K2.6, GLM-5.2 and GLM-5.1, five Qwen models, Gemma 4 31B, Mistral Nemo and NVIDIA Nemotron 3 Nano Omni. Quantization varies: fp8 for most, fp4 for Gemma 4 31B and GLM-5.2, int4 for Kimi K2.6 and mxfp4 for Kimi K3. There are no GPT, Claude or Gemini models. The catalog was much larger before 2026; Chutes claimed 150+ models in November 2025.
You pay by the token or take a plan: Plus at 10 USD a month for 2,000 requests a day, or Pro at 20 USD for 5,000. Plans are capped at 50 and 100 USD of pay-as-you-go value a month, so the daily count is often out of reach, and the terms keep them to "casual, direct usage": automated or production use must be pay-as-you-go.
Through 2025 Chutes was the free backend for DeepSeek on OpenRouter and in roleplay apps such as Janitor AI. That ended in stages. The Early Access tier of 200 free requests a day lost TEE models on February 27, 2026 and closed on March 15. The pricing page now says: "We do not offer a free tier at this time." A 3 USD Base plan that existed into at least May 2026 is gone. Robo Rhythms, an affiliate site, reports DeepSeek V3.2 rising from 0.24 / 0.42 USD to 1.00 / 1.00 USD per million tokens without notice, and the current list shows 1.00 / 1.00 USD. Chutes' own March 2026 post says revenue per million tokens rose 37.7% after February 1.
Chutes hosts the models itself and sets its own prices; the pricing page says "No subscription, no minimum, no markup."
| Model, per million input / output tokens | Chutes | Maker's own price | Rank on OpenRouter |
|---|---|---|---|
| Kimi K2.6 | 0.45 / 1.95 USD | 0.95 / 4.00 USD, Moonshot AI (OpenRouter) | 2nd cheapest of 17 hosts |
| GLM-5.1 | 0.98 / 3.08 USD | 1.40 / 4.40 USD, Z.AI (OpenRouter) | 3rd cheapest of 13 hosts |
| DeepSeek V4 Flash 0731 | 0.44 / 1.32 USD | DeepSeek sells the newer V4.1 Flash at 0.30 / 1.20 USD peak (DeepSeek) | not compared |
Chutes' prices come from its pricing page. It charges about half Moonshot's price for Kimi K2.6 and about 30% less than Z.AI for GLM-5.1, but its older DeepSeek V4 Flash costs more than DeepSeek's own newer Flash model.
The low prices are subsidised. Pine Analytics estimated in March 2026 that subnet 64 received about 518 TAO a day in emissions, roughly 52 million USD a year, against 1.3 to 2.4 million USD of customer revenue: 22 to 40 times more. Chutes' share of emissions was about 14.39% then; readings cited by Own Your Mind put it at 5.82% on September 22, 2026. A DefiLlama contributor noted that Chutes stopped publishing its revenue series after September 6, 2026.
No independent speed benchmark has been published: Chutes does not appear on 16 Artificial Analysis provider pages for models it serves or has served.
Hacker News is mixed. One commenter called about 5,000 prompts a day for 20 USD practically unlimited (May 2026); another spent 30 USD in an hour or two without prompt caching (August 2025). Arcanum RPGs, which says it has no commercial tie, writes that old "free Chutes DeepSeek" guides are outdated and that a busy model has no fallback. On GitHub, users of Chutes' e2ee-proxy report that its built-in TLS certificate expired on September 22, 2026. Chutes has no Trustpilot profile.
/plugin marketplace add Veightor/chutes-agent-toolkit, holds nine SKILL.md skills: setup, routing, TEE checks, usage and billing, sign-in, platform operations, deployment, MCP portability and agent registration. Six are marked beta. The repository belongs to a personal GitHub account, not Chutes' chutesai organization.llms.txt.Chutes' privacy policy (March 16, 2026) says it does not log or store prompt or response content for public LLM inference and does not train on user data; in standard mode servers may read request data in memory. Chutes' own chat apps store conversations. No SOC 2 or ISO claim was found. The terms follow Nevis law with binding arbitration and a class-action waiver. Crypto payments are final, Stripe refunds happen only in limited cases at Chutes' discretion, and any balance may be forfeited if an account is ended. status.chutes.ai watches only the website and an API heartbeat, so it shows 100% even when models are at capacity.
Skip Chutes for production apps, for closed models, and wherever you need guaranteed capacity, a fallback host or a company accountable under US or EU law. It suits hobby and roleplay users who want cheap open models on a flat plan and accept occasional capacity errors.
Sources
Changelog
APIMODELS is a pay-as-you-go AI API relay that sells about 150 models through one key at its own prices, mostly below the makers' list prices, and spends the margin on backup channels: 98.62% of requests succeeded over 30 days, moderation rejections excluded. APIMODELS is our own product: it is made by the team behind GenGrades.
ZenMux is an LLM gateway, operated under its terms by NexaMind Singapore Pte. Ltd., that sells Claude, GPT, Gemini and many Chinese models through OpenAI, Anthropic and Vertex-style endpoints at the makers' list token prices. It adds automatic credits for bad or slow answers, which it calls insurance, and 20 to 200 USD monthly coding plans.
Vercel AI Gateway is Vercel's hosted API for about 480 text, image, video and speech models behind one key, billed from prepaid credits at the providers' list prices with no markup, including when you bring your own keys. It became generally available on August 21, 2025, and gives each Vercel team 5 USD of free credit a month.
Requesty is a London-based LLM gateway, run by Requesty Ltd and seed-funded by 20VC in 2025, that sells OpenAI- and Anthropic-compatible access to more than 600 models at list price plus a 5% top-up fee. It stands out for routing policies, spend controls and an EU endpoint in Frankfurt.
Poe API is Quora's developer API for the bots in its Poe chat app, opened on July 31, 2025, with OpenAI-compatible endpoints and an Anthropic-compatible one for official Claude bots. Calls spend the same points as a Poe subscription, so there is no separate API bill; current Claude bots cost about 1% more than Anthropic's list price.
OpenRouter is a New York company's API relay that sells access to more than 680 models from over 100 hosting providers through one key, at each provider's list price plus a 5.5% fee on credit purchases. Stripe agreed to buy it on August 19, 2026; no closing had been announced by October 21, 2026.