Chutes
Listed at https://chutes.ai
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | Per-token pay-as-you-go, quoted per 1M input tokens (verified on chutes.ai/api/models 2026-09-06): Qwen3.5-397B-A17B $0.45, DeepSeek-V3.2 $1.00, DeepSeek-V4-Flash $0.44, Kimi-K2.6 $0.58, GLM-5.2 $1.25, Qwen3-32B $0.104, Gemma-4-31B $0.12, Qwen3.6-27B $0.30, Qwen3.8-27B $0.32 | per million tokens |
| Output | Per 1M output: Qwen3.5-397B $3.00, DeepSeek-V3.2 $1.00, DeepSeek-V4-Flash $1.32, Kimi-K2.6 $3.40, GLM-5.2 $3.95, Qwen3-32B $0.416, Gemma-4-31B $0.37, Qwen3.8-27B $2.50 | per million tokens |
🤖 Supported Models (33)
✨ Pros
- ✓Decentralized open-weights inference network that hosts SOTA open models minutes after release (Kimi K2.6, GLM 5.2, Gemma 4 31B, DeepSeek V4 Flash all live)
- ✓Fully transparent, no-markup per-token public pricing — the 'priced in the open' aggregator-side contrast to OpenRouter / Kilo / FetchAI
- ✓All served models run on confidential TEE compute for privacy and isolation
- ✓Very low entry price points: Qwen3-32B $0.104/$0.416, Gemma-4-31B $0.12/$0.37 per 1M
- ✓Private Chutes: deploy your own code on a TEE GPU billed per second, from $1.80/hr (RTX Pro 6000 96GB)
- ✓OpenAI-compatible endpoint at https://llm.chutes.ai/v1 plus official Python/CLI — low migration friction
⚠️ Cons
- ×No free tier — you must top up the app balance before first call (unlike Novita/DeepInfra which offer small free credits)
- ×No mainland-China endpoint or payment method; production traffic from China needs a proxy
- ×Decentralized / third-party miner GPU topology means weaker regional determinism than owned-datacenter providers; latency varies with node placement
- ×No first-party frontier proprietary models — focused on the open-weights ecosystem (no direct GPT-6 Astra / Claude / Gemini)
- ×Catalog surfaces filter to 'hot' models first; the full model/context table is pulled from the live API rather than one static list
🎯 Best For
Developers and agent teams that want the latest open-weights models at very low per-million prices through one OpenAI-compatible API, with confidential TEE compute as a privacy upside; anyone who needs per-second, elastic custom-code or fine-tuned inference on a dedicated TEE GPU; and teams that prefer a transparent no-markup open-source-native marketplace over a closed resale aggregator.
💰 Pricing & Plans
| Model | Input ($/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; strong 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; flagship Kimi on the network |
| zai-org/GLM-5.2-TEE | $1.25 | $3.95 | frontier agentic/coding open model |
| Qwen3-32B-TEE | $0.104 | $0.416 | 41K; cheap small workhorse |
| google/gemma-4-31B-turbo-TEE | $0.12 | $0.37 | 131K; budget edge model |
| Private Chute (TEE GPU) | from $1.80/hr | per second | RTX Pro 6000 96GB; +1x deploy fee (3x hourly) |
🔧 API & Developer Experience
- •OpenAI-Compatible Base URL: Drop-in OpenAI-compatible endpoint at https://llm.chutes.ai/v1 (chat via /v1/chat/completions). Existing OpenAI SDKs and curl work by swapping base_url and using a CHUTES_API_KEY bearer token.
- •Official Tooling: First-party 'chutes' CLI (Python/MIT, github.com/chutesai) plus docs.chutes.ai; create API keys in the app or CLI and deploy private chutes with a single command.
- •Live Pricing API: https://chutes.ai/api/models returns current per-1M input/output rates for every model in one JSON call — costs are pulled live so the billed rate always matches the listed rate.
- •No Free Tier / Top-up Model: No free tier; top up the app balance (25+ Stripe methods incl. cards, crypto, bank, plus TAO/Bittensor alpha tokens) then pay-as-you-go per token with no minimum.
- •Optional Subscriptions: Plus $10/mo (6% off beyond a daily quota) and Pro $20/mo (10% off) bundle a predictable monthly budget; Enterprise offers volume discounts and custom rate limits.
- •Private Chutes (Serverless TEE): Deploy your own code or a non-catalog model on a confidential TEE GPU (RTX Pro 6000 96GB from $1.80/hr) billed per second, isolated by default; idle instances auto-shutdown.
⚡ Why Decentralized: No-Markup Open Inference
Chutes is the open-source, decentralized inference network that competes with resale aggregators on a different economic model: it hosts the open-weights models itself. Models run on a Bittensor-driven network of independent GPU miners inside confidential TEE compute, and pricing is in the open — per-1M token rates carry no markup and are pulled live from the pricing API. That is why the newest SOTA open models appear minutes after release, and why a frontier 397B MoE like Qwen3.5 sits at $0.45-in/$3.00-out while small models run at ~$0.10/$0.42. Where OpenRouter is a neutral resale shop, Chutes is a vertically-integrated, open-source-native marketplace behind the same OpenAI-compatible endpoint. The flip side of decentralization: no single owned-datacenter region to pin latency to, and no proprietary frontier models (GPT-6 Astra / Claude) — you get the open-weights frontier, fast and transparent.
🌐 Regional Availability & Latency
Chutes is a decentralized network: inference flows through a single global OpenAI-compatible endpoint (https://llm.chutes.ai) over a Bittensor network of independent GPU miners, so latency follows miner node placement rather than an owned regional data-center map. North America and Europe generally see low latency on the listed TEE instances; Asia-Pacific round-trips vary with where the winning miner GPU 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, so mainland production traffic should route through a proxy or a relay. For privacy-conscious teams outside China, confidential-TEE execution is a genuine differentiator; for latency-critical China-facing traffic, pair a domestic provider (Qwen, Zhipu, Doubao) as the China layer.