Per-token rates. No minimum on the API.
Public rate card for the Inference API, Dedicated Deployments, and Managed Fine-Tuning. No "contact sales for pricing" gates on any standard SKU. Custom pricing exists for Dedicated 32-GPU+ deployments — the rate card stays public, sales gets called for the discount.
Per million tokens, input and output priced separately.
Output is priced 2–4× input because output is the GPU-bound side of inference. All rates are list — no closed-lab discount columns, no margin math, no per-visitor pricing games.
| Model | Architecture | Input / 1M | Output / 1M |
|---|---|---|---|
| Llama-3.1-70B-Instruct | dense · 70B | $0.60 | $0.80 |
| Llama-3.1-70B FP8 | dense · 70B · quantized | $0.35 | $0.45 |
| Llama-3.1-405B-Instruct | dense · 405B | $3.00 | $5.00 |
| DeepSeek-V3 | MoE · 671B / 37B active | $0.30 | $1.10 |
| Mixtral-8x22B-Instruct | MoE · 141B / 39B active | $1.20 | $1.20 |
| Hermes-3-405B | dense · 405B · fine-tune of Llama | $3.00 | $5.00 |
Pricing principle: frontier sizes carry frontier prices. We are not the cheapest. We are positioned mid-market — cheaper than closed-lab frontier by 2–4×, near parity with other open-model inference clouds, with region-residency and contracts they don't offer.
Single-tenant nodes and clusters. Predictable monthly billing.
Minimum 3-month commit. Annual commits discount ~20%. No usage overage — the whole point of dedicated is predictable pricing, the node is yours for the term.
| Configuration | GPUs | 3-month / mo | 12-month / mo |
|---|---|---|---|
| 1× B200 node | 8× B200 SXM5 | $5,500 | $4,500 |
| 1× B300 node | 8× B300 | $6,200 | $5,000 |
| 2× B200 cluster | 16× B200 · InfiniBand | $10,500 | $8,500 |
| 4× B200 cluster | 32× B200 · InfiniBand | $20,000 | $16,000 |
Multi-region replication: same deployment in two regions bills at 1.9× single-region (10% off the second region). Larger clusters than the rate card supports — 8×, 16×, 32× nodes — get custom pricing via sales.
LoRA on the same models you serve from the API.
Fixed setup fee + per-million-token training charge. No surcharge to serve the adapter afterwards — adapter inference bills at the base-model API rate. Up to 10 adapters free per paying account; beyond that, $5/month per adapter.
| Base model size | Per training run | Typical runtime |
|---|---|---|
| 7B – 13B base | $50 + $0.20 / M tokens | 4 – 16 hours |
| 70B base | $200 + $0.40 / M tokens | 8 – 24 hours |
| 405B base | $1,000 + $1.00 / M tokens | 12 – 48 hours |
Full fine-tuning (non-LoRA) is Phase 2 and quoted custom. Adapter storage is near-free because the point is the adapter lives with us — your weights, our infra, no egress.
Pricing FAQ.
How does the free tier work?
Every new account gets $10 of API credit on signup, hard rate-limited to 1 RPS until you attach a billing method. Past the credit, you pay per token at the rates above. No tier on Dedicated or Fine-Tuning.
Is there a minimum commit on the API?
No. The Inference API is fully metered — you pay per token, no monthly minimum, no overage tier. Dedicated has a 3-month minimum and gets a discount at 12 months.
Can I get an annual discount?
On Dedicated, yes — ~20% off the 3-month rate. Inference API and Fine-Tuning are list price; we do not discount metered SKUs.
What currency do you bill in?
USD only at launch. EUR and INR invoicing lands in Q3, in the regions where the host stack is already region-resident.
What happens if I exceed my rate limit?
Soft throttle first — we slow you down before we 429 you. Sustained over-the-limit traffic returns 429 with a Retry-After header. Limits scale automatically with verified billing history; large customers raise them through Support.
How do you compare to OpenAI, Anthropic, Together?
Frontier sizes (Llama-405B, Hermes-405B) price 2–4× under closed-lab equivalents. Open-model peers (Together, Fireworks) price at parity — our edge there is the region-residency, the contracts, and the SOC 2 stack, not the raw $/token number.
$10 credit on signup. No card required.
Hard rate-limited to 1 RPS until you attach billing. That's enough to wire the API into your stack and decide whether to put traffic on it.