Let there be tokens.
Inference for token hungry workloads like long running agents, evals and batched workloads at a fraction of the cost.
Start self-serve in minutes with high-throughput async or 24-hour batch tiers.
Dedicated capacity (including real-time), custom models and high-volume workloads
Frontier-class output at a fraction of the cost
$51,900 USD with 90% of input served from cache
$51,180 USD with 90% of input served from cache
$11,663 USD with 90% of input served from cache
Save $46,600 per billion tokens at list price.
Intelligence score (Index v4.3) via Artificial Analysis. Cached figures assume a 90% hit rate at each model's cached-input rate per 1M: $0.22 Doubleword, $1.00 Astra, $0.20 Fable 5.1.
Scale smarter with latency-based pricing.
The same model at three speeds. Choose realtime, async or 24-hour batch, and pay less the longer you can wait.
Realtime, async and batch inference pricing
Low latency interactive model testing endpoints.
Need guaranteed, high-throughput capacity without rate limits? We deploy dedicated, optimised infrastructure around your workload, tuned for higher cache hit rates, throughput at scale and lower latency.
Talk to us about dedicated capacity →Pricing that scales with your workload.
Choose the model that best fits your business needs.
Self-Service Inference
Whenever you need it, at the speed and cost that best fits your workload.
Get started- Access to all major open-weight models
- Usage-based pricing with prepaid credits
- Prompt caching, tool calling and structured outputs
- High-throughput async and batch inference
- Zero Data Retention (ZDR) included
- Community and standard support
At-Scale Inference
For specialized workloads requiring dedicated engineering and white-glove support.
Talk to us- Everything available to self-service
- Dedicated infrastructure for guaranteed, high-throughput real-time capacity without rate limits
- Increased rate limits
- Custom models
- Workload-specific inference optimization (e.g., speculators) for improved throughput and reduced token costs
- Region-locked data processing (e.g., US or EU-based inference)
- Engineering support for workload optimization
- White-glove onboarding & Proof of Concept (PoC) support
- Volume pricing
- Signed MSA, DPA, and Custom SLAs
Let there be tokens.
Open-weight inference built for workloads that don't stop at a few thousand tokens.
Get started nowGet up to $50 in free credits with your first deposit.