Priced by how you use it.

What is the difference between the three tiers?

The tiers are priced by usage pattern, not seat count or key count. Human covers interactive use: IDE assistants, manual scripts, occasional API calls where a person decides when each request happens. Developer covers automation and pipelines: scheduled jobs, eval runs, batch processing initiated by a human but not requiring their presence. Agentic covers production systems running continuously without human initiation -- loops, pipelines, and workloads where the whole point is removing the human from the request path.

How does Agentic tier pricing work?

Agentic is provisioned capacity, not a flat rate. The monthly minimum is a retainer that reserves dedicated GPU infrastructure for your workloads. Usage scales with your agents up to an agreed ceiling. Your agents are never queued behind other customers' workloads. Accounts that consistently use significantly below their provisioned capacity are moved to the Developer tier -- the reservation only makes sense if the capacity is being used.

Why is there no flat rate for agent workloads?

A production agent running continuously can consume GPU capacity worth multiples of any reasonable flat monthly fee within days. Flat-rate pricing for agent workloads either caps the agent (defeating the purpose) or is not sustainable at the infrastructure cost. Provisioned capacity is honest about what it costs to run dedicated GPU infrastructure and gives you a guarantee in return.

Does Marigold train on my data?

No. Inference requests are not retained for training, fine-tuning, or any other purpose. Outputs are stored briefly for retrieval and then deleted. Agentic tier accounts can obtain a GDPR data processing agreement that sets this out contractually.

Where does my data go?

Inference runs on private AWS infrastructure in London (UK) by default. Data does not leave that region. No third-party model provider receives your inputs. Output references are purged after a configurable retention window.

What is the OpenAI-compatible endpoint?

POST /v1/chat/completions accepts the same request shape as the OpenAI Chat Completions API. Switch the base URL and API key; the model parameter maps to a Marigold model name. The developer page has a before-and-after example.

What is the difference between Marigold and self-hosting a model directly?

Self-hosting requires provisioning compute, managing model weights, implementing an inference API, and maintaining all of it. Marigold provides the typed async API, weight caching, and an eval surface as a managed layer on private AWS infrastructure. You control the deployment region and the model selection.