Pricing
Pay per token until reserved GPUs cost less.
Serverless bills each token on shared GPUs. Dedicated reserves replicas by the GPU-hour. Same API, same model IDs, so the switch is a billing decision, not a rewrite.
Serverless
from $0.008
per 1M tokens
Shared GPUs, billed per token. No minimums, $5 of credit to start.
Dedicated
from $2.49
per GPU-hour
H100, H200 and B200 replicas, billed by the second, with a 99.95% SLA.
Enterprise
Contract
committed volume
Invoiced billing, SSO and a security review, on either kind of capacity.
Calculator
What will a month cost?
Pick a model and a month of traffic. Dedicated is sized for peaks at 2× your average rate, so the replica count is what you would actually reserve.
Type a number, or use K, M or B (“250M”). Up to 100B each.
Estimated monthly cost
USD
Serverless
Cheaper$1,440.00
per month
Dedicated
$3,635.40
per month
Serverless $1,440.00, dedicated $3,635.40 a month. Serverless saves $2,195.40 a month at this volume.
Serverless saves $2,195.40 a month at this volume.
One replica pays for itself at 6.31B tokens a month, at your mix of 80% input ($0.58 per 1M blended).
2.5B tokens a month. Estimates before taxes and the free credit.
Serverless
Price per 1M tokens
Shared GPUs, billed per token, with input and output priced separately. The same prices as the catalog and every model page.
| Model | In / 1MInput / 1M | Out / 1MOutput / 1M |
|---|---|---|
| Llama 3.3 70B InstructMeta | $0.54 | $0.72 |
| Llama 3.1 8B InstructMeta | $0.10 | $0.14 |
| Llama 4 ScoutMeta | $0.18 | $0.59 |
| Llama 4 MaverickMeta | $0.27 | $0.85 |
| Qwen3 235B A22BAlibaba | $0.22 | $0.88 |
| Qwen3 32BAlibaba | $0.30 | $0.50 |
| Qwen3 8BAlibaba | $0.07 | $0.12 |
| Qwen2.5 Coder 32B InstructAlibaba | $0.28 | $0.44 |
| Qwen3 Coder 480B A35BAlibaba | $0.95 | $1.90 |
| DeepSeek V3.1DeepSeek | $0.55 | $1.65 |
| DeepSeek R1 0528DeepSeek | $0.70 | $2.40 |
| DeepSeek R1 Distill Llama 70BDeepSeek | $0.58 | $0.78 |
| Mistral Small 3.2 24BMistral AI | $0.10 | $0.25 |
| Mistral Large 3Mistral AI | $0.65 | $1.60 |
| Devstral Small 2Mistral AI | $0.09 | $0.30 |
| Gemma 3 27BGoogle | $0.20 | $0.40 |
| Phi-4 14BMicrosoft | $0.07 | $0.14 |
| Qwen2.5 VL 72B InstructAlibaba | $0.60 | $0.80 |
| BGE-M3BAAI | $0.01 | — |
| Nomic Embed Text v1.5Nomic | $0.008 | — |
| Qwen3 Embedding 8BAlibaba | $0.02 | — |
Images
Image models bill one price per generated image, whatever the prompt length.
| Image model | Size | Per image |
|---|---|---|
| FLUX.1 [dev]Black Forest Labs | 1024 × 1024 | $0.025 |
| FLUX.1 [schnell]Black Forest Labs | 1024 × 1024 | $0.003 |
| Stable Diffusion 3.5 LargeStability AI | 1024 × 1024 | $0.035 |
Dedicated
GPU-hours, billed by the second
Replicas reserved for you, with the same API and model IDs as serverless, and a 99.95% monthly uptime SLA. Models without a ready replica are set up on request.
| GPU | Memory | Per GPU-hour |
|---|---|---|
| NVIDIA H100 | 80 GB | $2.49/hr |
| NVIDIA H200 | 141 GB | $3.39/hr |
| NVIDIA B200 | 192 GB | $5.49/hr |
Ready replicas
One replica per row, always on for a month. Break-even is the monthly volume where one replica costs the same as serverless, at an even input/output mix; the calculator uses yours.
| Model | Per hour | The monthly volume where one dedicated replica costs the same as serverless, at half input and half output tokens. |
|---|---|---|
| Llama 3.3 70B InstructMeta | $4.98/hr | 5.77B tok/mo |
| Llama 3.1 8B InstructMeta | $2.49/hr | 15.1B tok/mo |
| Llama 4 ScoutMeta | $3.39/hr | 6.43B tok/mo |
| Qwen3 32BAlibaba | $2.49/hr | 4.54B tok/mo |
| Qwen2.5 Coder 32B InstructAlibaba | $2.49/hr | 5.05B tok/mo |
| DeepSeek R1 Distill Llama 70BDeepSeek | $4.98/hr | 5.35B tok/mo |
| Mistral Small 3.2 24BMistral AI | $2.49/hr | 10.4B tok/mo |
| Gemma 3 27BGoogle | $2.49/hr | 6.06B tok/mo |
Enterprise
Need a contract?
For teams that buy through procurement: the same capacity, on terms your legal and security reviews ask for.
- Pricing
- Committed-volume rates on serverless and dedicated
- Billing
- Invoiced monthly in USD
- SLA
- 99.95% monthly uptime, with service credits. Every dedicated endpoint is covered, self-serve or under contract.
- Identity
- SSO with SAML 2.0 and SCIM, with audit logs
- Compliance
- SOC 2 Type II, ISO 27001 and GDPR, with a DPA
- Support
- A named engineer and a shared Slack channel
FAQ
Before your first invoice
Serverless is billed per token. Input and output are priced separately per 1M tokens, and image models per image. Usage is added up across your keys and charged monthly, with no minimums and no seat fees. New accounts start with $5 of credit, no card required.
Dedicated replicas are billed by the second at the hourly rates above, from the moment a replica starts until you release it.
A new key starts at 600 requests a minute on serverless, and the limit rises as your account spends. Dedicated endpoints have no per-key request limit: throughput is set by the replicas you reserve. Each replica’s peak is listed in the table above.
Move when steady volume passes the break-even point for your model (the calculator works it out from your own input and output mix), or when you need reserved throughput and the 99.95% monthly uptime SLA. Your code does not change: the model ID, the request format and the SDKs stay the same, and the router sends your requests to replicas reserved for you instead of the shared pool. You can keep low-volume models on serverless and move only the busy ones.
Zero retention by default. Never used for training. Request logging is opt-in, kept 30 days.
Compliance: SOC 2 Type II, ISO 27001 and GDPR, with a DPA. A DPA with standard contractual clauses is available on every plan.
No. Serverless and dedicated are self-serve, with no minimum term: release a replica and its billing stops. The 99.95% SLA comes with dedicated either way: every dedicated endpoint is covered, self-serve or under contract. A contract adds committed-volume rates on serverless and dedicated, billing invoiced monthly in USD, SSO with SAML 2.0 and SCIM, and a security review with our reports under NDA. Talk to an engineer to start one.