Chutes API Review 2026: Decentralized Open-Weight Inference
APIRank already tracks the proprietary frontier — GPT-6 Astra, Claude Fable, Gemini — plus the big resale aggregators like OpenRouter. Today I am adding the other end of the open-weights market that had been missing from our catalog: Chutes, a decentralized inference network that hosts open-source models itself (not resells them) and runs every request on confidential TEE compute. It is genuinely active — the day I verified it, Kimi K2.6, GLM 5.2, Gemma 4 31B and DeepSeek V4 Flash were all live on the network — and its per-1M pricing is published in the open. This review covers the verified pricing, the OpenAI-compatible API, the decentralized/TEE angle, and where Chutes fits next to the aggregators we already compare.
TL;DR: Chutes (chutes.ai) is an open-source, decentralized inference network for open-weights models, served through a single global OpenAI-compatible endpoint at https://llm.chutes.ai/v1 and run on Bittensor-backed GPU miners inside confidential TEE compute. Verified first-party pricing (chutes.ai/api/models + /pricing FAQ, both fetched 2026-09-06): Qwen3.5-397B-A17B at $0.45 in / $3.00 out per 1M, DeepSeek-V3.2 at $1.00/$1.00, DeepSeek-V4-Flash at $0.44/$1.32 (1.0M context), Kimi-K2.6 at $0.58/$3.40, and small models like Qwen3-32B at just $0.104/$0.42. No markup, no free tier. Optional Plus $10/mo and Pro $20/mo plans. Private, per-second TEE GPU deployments from $1.80/hour. The trade-off: it is the open-weights frontier — no proprietary GPT-6 Astra / Claude — and there is no owned regional datacenter to pin latency to, nor any China endpoint or payment method.
Why Chutes, and Why Now
This week's APIRank briefings were dominated by the GPT-6 Astra launch and the Gemini 3.8 Flash family — both already covered on the site. What those stories made me realize is how thin our open-weights, non-resale coverage was. I list DeepInfra, Together, Novita and Hyperbolic as open-weights hosts, but I did not yet cover the decentralized segment where the models run on an independent, Bittensor-driven GPU network rather than one vendor's datacenter. Chutes is the clearest operator in that niche that is still actively shipping: the GitHub org shows commits through early September 2026, and its consumer products (Chutes Chat and Chutes Search) get real use. It also publishes its whole live catalog and pricing as JSON, which makes verification uncommonly easy.
Worth a caveat up front: Chutes compares itself to OpenRouter, Kilo, FetchAI and Phala on its own homepage (the same row of logos that sit above its hero). I treat that as a feature, not a knock — it means the decentralized tier is positioning as an aggregator alternative, which is exactly the purchasing decision a developer with a lot of open-weights traffic faces today.
Verified Pricing (2026-09-06)
Chutes is aggressively transparent about cost. The model catalog and every per-1M input/output rate are served as live JSON at https://chutes.ai/api/models, so the number on the pricing page is the number you are billed. Below is a snapshot I verified directly from that API on 2026-09-06:
| Model | Input $/M | Output $/M | Notes |
|---|---|---|---|
| Qwen3.5-397B-A17B-TEE | $0.45 | $3.00 | 262K ctx; frontier open reasoning MoE |
| DeepSeek-V3.2-TEE | $1.00 | $1.00 | 262K ctx; popular open chat model |
| DeepSeek-V4-Flash-0731-TEE | $0.44 | $1.32 | 1.0M ctx; low-cost Flash class |
| moonshotai/Kimi-K2.6-TEE | $0.58 | $3.40 | 262K ctx; flagship Kimi line |
| zai-org/GLM-5.2-TEE | $1.25 | $3.95 | frontier agentic/coding open model |
| Qwen3-32B-TEE | $0.104 | $0.416 | 41K ctx; cheap small workhorse |
| google/gemma-4-31B-turbo-TEE | $0.12 | $0.37 | 131K ctx; budget edge model |
Table note: all rates and context windows verified from the live chutes.ai/api/models pricing API on 2026-09-06.
The pricing model itself is simple and honest-sounding: no markup tiers, no middlemen, pay for the tokens you actually consume. If you do not want to top up as you go, the optional Plus ($10/mo) and Pro ($20/mo) plans bundle a daily request quota at 6% and 10% off the pay-as-you-go rates respectively. There is no free tier — the official FAQ says plainly: "We do not offer a free tier at this time. Top up your account in the app to get started." That is a real difference from aggregators like Novita or DeepInfra that hand out small free credits at signup, and it is the single biggest onboarding complaint a new user will hit.
API Surface and Developer Experience
Migration friction is minimal because Chutes is OpenAI-compatible. The base URL is https://llm.chutes.ai/v1 and chat goes to /v1/chat/completions; you authenticate with Authorization: Bearer $CHUTES_API_KEY. An existing OpenAI SDK works after swapping the base URL and key, and the official home-page snippet shows a plain requests call in Python. Chutes also ships a first-party chutes CLI (Python, MIT, under the chutesai GitHub org), which handles API keys, model queries, and — its signature feature — one-command deploys of a Private Chute.
The Private Chute is the second half of the product and deserves its own mention. If a model is not in the public catalog, or you need to run your own code or a private fine-tune with full isolation, you deploy it onto a confidential TEE GPU. The self-serve class Chutes currently lists is the RTX Pro 6000 (96 GB Blackwell) at $1.80 per hour, billed per second, with a one-time deployment fee equal to 3× the hourly rate ($5.40). Idle instances shut down automatically, so no GPU-seconds are wasted. This is closer to a leaner, TEE-flavored RunPod than to a pure model API — useful if you have custom weights but do not want to stand up a full GPU cluster.
Chutes vs the Aggregators
The cleanest way to think about Chutes is as the vertically-integrated, open-source-native alternative to a resale aggregator. OpenRouter, Kilo and their peers are neutral shops that route to whatever any upstream provider lists — they give you one key and a huge model menu, but they add a routing layer and their inventory is someone else's. Chutes instead owns the stack: it hosts the open-weights models on its own Bittensor-driven network, isolates each in a TEE, and publishes no-markup rates. Its homepage even names its own competitors — "OpenRouter, Kilo, FetchAI, Phala, Pax Historia" — which makes the positioning unambiguous.
For an APIRank audience the practical comparison comes down to a few axes. Price: a frontier 397B open MoE like Qwen3.5 at $0.45 in/$3.00 out is aggressive against first-party proprietary pricing, and the tiny models are astonishingly cheap. Model freshness: Chutes advertises SOTA open models "minutes after release," and the live catalog confirmed it — the newest Kimi, GLM and DeepSeek iterations were all present the day I checked. Privacy: confidential TEE execution is a genuine differentiator over multi-tenant clouds. The downsides: no proprietary frontier access (you will not call GPT-6 Astra or Claude through it), no free tier, and the decentralized topology means you cannot point at a single owned region for latency the way you can with Hyperbolic's or Together's own capacity.
China Access and Latency
Chutes serves everything through one global OpenAI-compatible endpoint in front of a decentralized GPU network, so latency follows miner node placement rather than a published regional datacenter map. North America and Europe generally see low latency on the listed TEE instances; Asia-Pacific round-trips vary with where a request is scheduled. There is no mainland-China regional endpoint, no China payment method (native on-ramps are Stripe and TAO/crypto), and no China data-residency option. Mainland developers should route through a proxy or a relay that fronts the Chutes API, and for any user-facing, latency- or residency-sensitive surface serving China traffic, pair a domestic provider (Qwen, Zhipu, Doubao) as the China-facing layer — the same guidance I give for every overseas open-weights host in the catalog.
Who Should Use Chutes?
- Open-weights-first teams who want a single OpenAI-compatible key across the newest Kimi/GLM/Qwen/DeepSeek releases, and who like that the pricing is transparent and markup-free.
- Cost-sensitive, high-volume workloads — the small-model tier (Qwen3-32B at ~$0.10 in, Gemma-4-31B at ~$0.12 in) is genuinely hard to beat per token.
- Privacy-conscious developers who value confidential TEE execution and isolation over a shared multi-tenant cloud.
- Anyone running custom weights or code who wants per-second TEE GPU capacity without managing a cluster (the Private Chute path).
It suits less well the team that needs proprietary frontier access, a free tier to prototype on, or a fixed, owned-region latency profile. And note the model catalog can shift — Chutes filters its homepage to "hot" models first, so treat the live API as the source of truth for what is actually servable on any given day.
Chutes FAQ
What is Chutes in 2026?
Chutes is an open-source, decentralized inference network for open-weights models, served through a single global OpenAI-compatible endpoint (https://llm.chutes.ai/v1) and run on a Bittensor-driven network of independent GPU miners inside confidential TEE compute. It hosts open models itself rather than reselling another vendor's listings, publishes no-markup per-1M pricing, and also offers per-second Private Chute GPU deployments.
How much does the Chutes API cost?
Pay-as-you-go per 1M tokens, no markup. Verified 2026-09-06: Qwen3.5-397B-A17B $0.45/$3.00 (262K ctx), DeepSeek-V3.2 $1.00/$1.00, DeepSeek-V4-Flash $0.44/$1.32 (1.0M ctx), Kimi-K2.6 $0.58/$3.40, GLM-5.2 $1.25/$3.95, Qwen3-32B $0.104/$0.416, Gemma-4-31B $0.12/$0.37. Plus $10/mo and Pro $20/mo add a discounted daily quota. No free tier.
Is the Chutes API OpenAI-compatible?
Yes. Endpoint https://llm.chutes.ai/v1, chat at /v1/chat/completions, auth via CHUTES_API_KEY bearer token — OpenAI SDKs work after swapping the base URL. A first-party 'chutes' CLI (Python, MIT) handles keys, model queries and one-command Private Chute deployments.
Can I use Chutes from mainland China?
Not directly — no mainland endpoint, no China payment method, no CN data-residency option. Mainland production traffic needs a proxy or relay. For latency- or residency-sensitive China-facing traffic, a domestic provider (Qwen, Zhipu, Doubao) is recommended as the China layer.
Does Chutes host proprietary frontier models?
No. The catalog is open-weights only — Qwen, DeepSeek, Moonshot Kimi, Zhipu GLM, Google Gemma, and others — with no GPT-6 Astra / Claude / Gemini access. For proprietary frontier models you still go first-party or to a resale aggregator.