Subscription plans with 5-hour and weekly usage budgets and wallet spillover — plus per-model API rates synced live with our catalog for pay-as-you-go usage.
One price, a generous usage budget per window, and a dashboard that shows exactly what is left. Every plan runs on recycled GPUs in regions with access to renewable energy.
Every number on this page is the binding one. The same budgets, limits and plan rules appear in Section 18 of the Terms of Service, and there is no hidden fair-use quota.
For one person and their agent. Every model in the catalog, a 5-hour usage budget that resets all at once, and a weekly budget that keeps a runaway loop from eating the month. Budgets are measured in dollars of inference at our published rates, so a small model or a cached prompt barely dents them.
$1.40 of usage per 5-hour window, $14 per week
Every model in the catalog (27)
1 request at a time
128K context
10M embedding tokens / mo
Up to 3 active API keys
For daily agent work. Two and a half times the Solo budget, three parallel requests so tool calls do not queue behind each other, and 256K context for repo-scale prompts. The plan most people building with opencode, Hermes or Cursor end up on.
$3.50 of usage per 5-hour window, $35 per week
Every model in the catalog (27)
3 parallel requests
256K context
50M embedding tokens / mo
Up to 5 active API keys
For agents that never log off. Six times the Solo budget, six parallel requests, routed first when the fleet is busy. Built for always-on assistants and multi-agent pipelines that run while you sleep. If you will use more than this, talk to us about dedicated capacity.
$8.40 of usage per 5-hour window, $84 per week
Every model in the catalog (27)
6 parallel requests, routed first when the fleet is busy
256K context
150M embedding tokens / mo
Up to 10 active API keys
No subscription — fund a wallet, pay published rates per token. Credit you pay for never expires. Bonus credit on larger packs is spent first and expires after 12 months.
Top-up presets
Custom amounts $5–$500
Recycled GPUs
Built from reclaimed GPUs and server hardware — extending useful life instead of manufacturing new silicon.
Renewable-ready
Deployed in regions with access to renewable energy.
No evaporative cooling
Designed to avoid evaporative water cooling.
Pay-as-you-go token, page, and endpoint pricing — synced live with our catalog.
| Model | Provider | Input / 1M tok | Output / 1M tok | Cached / 1M tok | Pages / 1K | Endpoint |
|---|---|---|---|---|---|---|
Qwen3.8 27BFLAGSHIP | Alibaba Cloud | $0.35/1M | $2.75/1M | $0.0875/1M | — | /v1/chat/completions |
DeepSeek V4 FlashFEATURED | DeepSeek | $0.191/1M | $0.506/1M | $0.0478/1M | — | /v1/chat/completions |
Qwen3.8 Flash Next | Alibaba Cloud | $0.15/1M | $0.47/1M | $0.0375/1M | — | /v1/chat/completions |
Qwen3.6 35B A3B | Alibaba Cloud | $0.15/1M | $1/1M | $0.0375/1M | — | /v1/chat/completions |
Qwen3.5 9B | Alibaba Cloud | $0.1/1M | $0.15/1M | $0.025/1M | — | /v1/chat/completions |
Gemma4 31B | Google DeepMind | $0.12/1M | $0.37/1M | $0.03/1M | — | /v1/chat/completions |
Gemma4 26B A4B | Google DeepMind | $0.06/1M | $0.33/1M | $0.015/1M | — | /v1/chat/completions |
Gemma4 12B | Google DeepMind | $0.05/1M | $0.15/1M | $0.0125/1M | — | /v1/chat/completions |
Gemma4 E4B | Google DeepMind | $0.06/1M | $0.12/1M | $0.015/1M | — | /v1/chat/completions |
Gemma4 E2B | Google DeepMind | $0.05/1M | $0.1/1M | $0.0125/1M | — | /v1/chat/completions |
Ornith 1.5 35B A3B | Ornith AI | $0.07/1M | $0.7/1M | $0.0175/1M | — | /v1/chat/completions |
Ornith 1.5 9B | Ornith AI | $0.1/1M | $0.15/1M | $0.025/1M | — | /v1/chat/completions |
Laguna S 2.1 | Poolside | $0.1/1M | $0.2/1M | $0.025/1M | — | /v1/chat/completions |
MiMo V2.6 Distill 9B | Xiaomi | $0.1/1M | $0.15/1M | $0.025/1M | — | /v1/chat/completions |
Ling 3.0 Tiny | inclusionAI | $0.02/1M | $0.11/1M | $0.005/1M | — | /v1/chat/completions |
LFM2.5 8B A1B | Liquid AI | — | — | — | — | /v1/chat/completions |
Granite 4.1 8B | IBM | — | — | — | — | /v1/chat/completions |
Granite 4.0 H Tiny | IBM | — | — | — | — | /v1/chat/completions |
Mistral Nemo 12B | Mistral AI | $0.02/1M | $0.03/1M | $0.005/1M | — | /v1/chat/completions |
GPT-OSS 20B | OpenAI | $0.029/1M | $0.14/1M | $0.0073/1M | — | /v1/chat/completions |
Muse Glimmer 30B | Unsloth | — | — | — | — | /v1/chat/completions |
Hermes Compressor | Cobble Labs | $0.06/1M | $0.33/1M | $0.015/1M | — | /v1/chat/completions |
| Model | Provider | Input / 1M tok | Output / 1M tok | Cached / 1M tok | Pages / 1K | Endpoint |
|---|---|---|---|---|---|---|
GLM-OCR | Zhipu AI | — | — | — | $0.08 | /v1/ocr |
DeepSeek OCR2 | DeepSeek | — | — | — | $0.08 | /v1/ocr |
| Model | Provider | Input / 1M tok | Output / 1M tok | Cached / 1M tok | Pages / 1K | Endpoint |
|---|---|---|---|---|---|---|
EmbeddingGemma 300MFEATURED | Google DeepMind | $0.01/1M | — | — | — | /v1/embeddings |
Fund your wallet for spillover and pay-as-you-go usage. Credit you pay for never expires. Larger packs include bonus credit, which is spent first and expires after 12 months.
Each plan includes a budget of inference per 5-hour window and per week, measured in dollars at the per-model rates below. Your first request starts the 5-hour window; when it ends, the whole budget comes back at once. The weekly window works the same way over 7 days, and a request must fit in both. Your dashboard shows what is left in each and when it resets.
Requests get a 429 that names the window and its reset time, and nothing is charged. If you opt into wallet spillover, usage past your budget is billed from your wallet at published catalog rates up to a monthly cap you set. Wallet credit you pay for never expires; bonus credit on larger packs expires after 12 months.
Every request is priced from its tokens at the per-model rates below: uncached input, cached input at a quarter of the input rate, and output. Plans burn that amount from your budget; wallet usage is charged it. A request the fleet declines for capacity is never charged and never touches your budget.
Yes, from billing settings in your account. Upgrades take effect immediately and your usage so far in the current windows carries over to the new budgets. Downgrades take effect at the start of your next billing cycle. If your subscription lapses, your API keys keep working on wallet billing at catalog rates until you subscribe again.
Each plan caps how many requests can run at once across all of your keys: 1 on Solo, 3 on Pro, 6 on Max. Extra requests queue until one finishes. Higher plans are also routed first when the fleet is busy, which means being refused last under load, not faster responses.
No. A self-service plan is for one person and their own agents and apps. Sharing keys across a team, reselling plan capacity, or opening a second account to stack budgets is prohibited and leads to suspension. Teams that need shared capacity should contact us about team or dedicated options.
Nothing is kept. Cobble does not store prompts or completions for any request, on any plan, and never trains on them. We keep request metadata only (model, token counts, cost, status, timing) for billing and abuse detection. All inference runs on hardware we operate in the United States.
Section 18 of the Terms of Service describes plans, budgets, windows, the wallet and plan changes in full; Part B of the Acceptable Use Policy covers fair use; the Privacy Policy covers data handling. The numbers on this page come from the same registry the API enforces.
Questions? Documentation · Terms of Service · Acceptable Use · Privacy
