Pricing
Two ways to pay. Both honest.
Ultra-fast is 1,000+ tps on reserved capacity, priced per workload. Standard is public per-token rates on every model in the catalog, self-serve. No minimums, no commits, no overage charges either way. $5 in starter credits to try Standard, no card.
Ultra-fast
1,000+ tps. Same models.
1,000+ tps, about 5x typical fast tiers. Reserved capacity on AWS Trainium and NVIDIA Blackwell, dedicated auto-scaling, contractual TTFT. Priced per workload, not per token, so we right-size the cluster to what your traffic actually looks like.
- Throughput
- 1,000+ tps
- TTFT
- < 250ms
- Capacity
- Reserved
- Cluster
- Single-model
Standard tier · 70 tok/s
Per-token rates.
Every model in the catalog, published rate per million tokens. Same API as ultra-fast, just without the reserved capacity. Self-serve, no contract.
| Model | $/1M input | $/1M cached | $/1M output |
|---|---|---|---|
GPT-OSS 120B OpenAI · 120B (MoE, ~5B active) | Contact sales | ||
Kimi K2.6 Moonshot · 1T MoE (~32B active) | $0.66 | $0.144 | $3.41 |
Qwen3 235B Alibaba · 235B MoE (~22B active) | Contact sales | ||
Kimi K2.6: Fast Moonshot · 1T MoE (~32B active) | $0.68 | $0.144 | $3.41 |
Kimi K2.7 Code Moonshot · 1T MoE (~32B active) | Contact sales | ||
GLM-5.2 Zhipu · 744B (MoE, ~40B active) | $1.20 | $0.2 | $3.00 |
Calculator
Math we don't hide.
Plug in monthly volume to see what your bill actually looks like. Type any number, and the slider scales exponentially up to 100B tokens.
Enterprise deployment
Your VPC. Our cluster. One bill.
Single-tenant deployment on AWS Trainium and NVIDIA Blackwell. Zero retention. SOC 2 Type II via Decart AI. VPC peering, contractual P99 SLAs, BAA on the roadmap. The full pitch lives at /enterprise.
- ▸Single-tenant on Trainium / Blackwell
- ▸SOC 2 Type II (via Decart AI)
- ▸GDPR + EU data residency
- ▸VPC peering, zero public egress
- ▸P99 latency SLA with credits
How billing works
Math we don't hide.
Each rule is tagged with the tier it applies to. Standard is per-token. Ultra-fast is per workload. Anything tagged Both holds either way.
- Standard: tokens, not minutes.
- We bill prompt + completion tokens reported by the upstream. No charge for capacity sitting idle between requests.
- Ultra-fast: priced per workload.
- Reserved capacity, sized to your traffic. One predictable monthly figure, no per-request surprises, but also not per-token economics.
- Standard: cached input is cheaper.
- When the upstream KV cache hits, those input tokens bill at the per-model cache rate (visible in the table above).
- Fine-tunes price the same.
- Your weights, our serving. No surcharge for the privilege: same per-token rate (Standard) or same cluster shape (ultra-fast).
- Hard spend caps. Per org, per key.
- Set a ceiling once. We stop accepting requests when you hit it and email you. Standard caps are dollar-denominated; ultra-fast caps are RPM-denominated.
- Credit doesn't expire.
- $1 of credit is $1 of usage, for the life of the account. No usage tiers, no bonus credit that vanishes.