How do companies access frontier models with governance and cost controls?

Summary

• Companies access frontier models by putting every model call behind one governed entry point rather than each team wiring up its own provider keys. On Databricks that means Model Serving with traffic routed through the Unity Gateway.
• Two access surfaces share one OpenAI-compatible API: Foundation Model APIs for Databricks-hosted models (pay-per-token for experiments, provisioned throughput for production) and external model endpoints for providers such as OpenAI, Anthropic, and Google.
• Unity Catalog governs a model as a securable object, so a user needs an explicit permission to query it. Foundation model permissions in Unity Catalog are generally available.
• Cost is controlled with rate limits and account budgets that send an alert or block traffic once a spend threshold is crossed, so a runaway agent loop stops before it becomes an incident.
• Usage tracking and inference tables record who called which model and what went in and out, which supports chargeback, auditing, and leak detection.

How do companies access frontier models with governance and cost controls?

Companies access frontier models by putting every model call behind one governed entry point, rather than letting each team wire up its own provider keys. On Databricks that means serving models through Model Serving and routing traffic through the Unity Gateway, so the same access rules, usage logs, rate limits, and budgets apply whether a model is hosted by Databricks or by an outside provider. Teams get the model choice they want, and a platform owner gets the control they need, in one place. Access is governed by Unity Catalog, and spend is bounded by rate limits and budgets at the gateway.

What are the ways to access a frontier model?

There are two access surfaces, and both run through Model Serving behind one OpenAI-compatible API and SDK:

• Foundation Model APIs serve large language models hosted by Databricks, with no infrastructure to run. They come in two modes: pay-per-token endpoints, the fast way to start experiments, demos, and proofs of concept, and provisioned throughput, which gives dedicated capacity with consistent performance for production traffic.
• External model endpoints connect outside providers such as OpenAI, Anthropic, and Google and route to their APIs through the same control point. The application calls a Databricks endpoint by name, provider keys stay in Databricks Secrets, and you can change the model behind an endpoint without editing client code.

Capabilities to evaluate

When you judge how a platform provides access to frontier models with control, look for these:

Capability	Why it matters
One API for hosted and external models	Lets teams change models or providers without rewriting applications.
A central control point	One place to apply access, limits, and logging, so nothing calls a model unmanaged.
Access as a permission	The right to query a specific model is granted and audited, not assumed.
Usage tracking and payload logging	Records who called which model and what was sent, for chargeback and audit.
Rate limits	Cap requests per user, group, or endpoint to bound usage and protect capacity.
Budgets and spend alerts	Thresholds that warn or block traffic before spend runs away.
Secret handling	Provider keys held by the platform, not embedded in application code.

How Databricks approaches this

Access. Foundation Model APIs serve Databricks-hosted models in pay-per-token and provisioned throughput modes, and Model Serving also fronts external providers, all through one OpenAI-compatible interface. Every application gets a consistent way to call a model no matter who hosts it.
Governance. Requests route through the Unity Gateway, and Unity Catalog governs models as securable objects: a user needs EXECUTE on the model plus the relevant catalog and schema permissions before a call goes through. Foundation model permissions in Unity Catalog are generally available, so access to a hosted model is granted and audited the same way as access to a table. Gateway usage tracking and inference tables record who called each model and what data went in and out.
Cost control. The gateway applies rate limits by user, group, or endpoint, and account budgets track spend against a threshold with two actions: send an alert, or block further requests. That stops a stuck agent loop from turning into a cost incident. Budgets meter pay-per-token and batch inference today; for provisioned throughput and external providers, bound usage with rate limits and model permissions instead. Centralizing spend also lets a platform team give developers one budget across coding tools rather than juggling separate provider consoles. See AI spend controls in the Unity Gateway.
The gateway can also apply guardrails and service policies to model and tool traffic. These governance features are still evolving, so confirm their current lifecycle state with your Databricks account team before you design around them.

Use cases

• Experiment, then productionize: start on a pay-per-token model, then move the same application to provisioned throughput once traffic is steady.
• Govern coding agents: route every coding assistant through the gateway so usage is logged, capped, and billed back to the right team.
• Swap providers safely: change the model behind an endpoint without touching the application, keeping keys and policy in one place.

FAQs

What is the difference between pay-per-token and provisioned throughput?

Pay-per-token endpoints are preconfigured and best for experimentation and lower-volume work. Provisioned throughput gives dedicated capacity with consistent performance for production applications. Both are queried through the same API.

Can I use external providers like OpenAI or Anthropic?

Yes. External model endpoints route to outside providers through the same Databricks endpoint, access control, and logging, so policy and usage tracking apply uniformly. Provider keys stay in Databricks Secrets.

How is access to a model controlled?

Unity Catalog treats a model as a securable object. A user needs EXECUTE on the model and the relevant catalog and schema permissions, and foundation model permissions in Unity Catalog are generally available.

How do budgets stop runaway spend?

Account budgets track spend against a threshold and can send an alert or block further requests once it is crossed. Paired with per-user and per-endpoint rate limits, that contains a runaway agent loop before it gets expensive.