# How do I choose an AI gateway?

> Choose an AI gateway by deciding who should own provider accounts, credentials, deployment, routing policy, billing and incident response. Then test the exact endpoint, model, streaming, tools, errors, usage and data policy your application needs. A long feature list is less useful than a short proof that the gateway fits your operating model and production payload.

_Last reviewed 2026-08-15 · [Editorial review](https://keeprouter.com/editorial-policy#editorial-team)_

## Start with the operating model

| Operating model | You buy or run | Typical reason to choose it |
|---|---|---|
| Managed model access | One gateway account, catalog and billing relationship | Fast setup and fewer provider accounts |
| BYOK control plane | Gateway policy around your provider credentials | Keep provider contracts while centralizing control |
| Self-hosted proxy | Software plus infrastructure, secrets and data stores | Infrastructure ownership and custom policy |
| Cloud AI platform | Models, identity, agents, evaluation and cloud services | Existing cloud governance and platform depth |
| Direct provider API | One provider contract and native API | Unique capability or the smallest possible path |

These products overlap but are not interchangeable. KeepRouter is managed model access with a public catalog and prepaid billing. LiteLLM is commonly self-hosted. Kong extends an API platform. Amazon Bedrock, Gemini Enterprise Agent Platform and Microsoft Foundry cover a much larger cloud platform surface.

## Write non-negotiable requirements

List the endpoint families, model capabilities, region or retention rules, authentication method, key scope, streaming semantics, tool behavior, output limits, spend controls, support expectations and deployment constraints. Mark each item as required, preferred or irrelevant. This prevents a demo feature from outweighing a missing production contract.

Use the [best AI gateways shortlist](/compare/best-ai-gateways) to select operating models, then open the product-specific comparison rather than treating the shortlist as a ranking.

## Test one representative path end to end

1. Create the narrowest key and budget possible.
2. Send a real non-streaming payload and record identity, model, route, usage and charge.
3. Test streaming through the terminal event.
4. Exercise tools, structured output or multimodal input if production needs them.
5. Trigger invalid auth, unsupported model, rate limit, timeout and upstream failure cases.
6. Inspect logs and confirm which sensitive fields are retained.
7. Compare the result with the old path using the same acceptance criteria.
8. Prove rollback before widening traffic.

The [gateway evaluation article](/blog/evaluate-ai-gateway) provides a fuller evidence matrix. For compatible clients, use the [OpenAI migration checklist](/blog/openai-compatible-api-migration-checklist) instead of assuming a base URL change proves parity.

## Calculate total ownership, not only token price

Include credit or platform fees, provider charges, cache behavior, failed attempts, logging plans, egress, infrastructure, on-call work and the cost of maintaining provider integrations. A managed service can cost more per unit and less in staff time. A self-hosted proxy can have no software fee and still require meaningful operations.

## Make the rejection reason explicit

A good decision record states why the chosen gateway fits and which requirement would force a different choice. That keeps the selection useful when models, pricing, routes or team constraints change. Review current official sources at purchase time because product terms and capabilities move quickly.

## Frequently asked questions

### What should I compare first?

Compare operating ownership first: provider accounts, credentials, deployment, routing, billing, logs and incident response.

### How many gateways should I test?

Test a small set representing the operating models you could realistically run. A large undifferentiated trial wastes time.

### Is model count a useful metric?

Only after filtering for the endpoints, regions, capabilities and commercial terms you need. Raw counts are rarely decision-ready.

### Should I choose the cheapest gateway?

Choose the lowest total cost that still meets the contract. Include staff time, failures, fees, infrastructure and migration work.

### How often should the decision be reviewed?

Review when a required model, route, price, data policy, team constraint or incident pattern changes.

## Sources reviewed

1. [KeepRouter OpenAPI](https://keeprouter.com/api/openapi.json)
2. [OpenAI API reference](https://developers.openai.com/api/reference/overview)
3. [Anthropic API documentation](https://docs.anthropic.com/en/api/overview)
4. [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework)

## Related guides

- [Best AI gateways by use case](https://keeprouter.com/compare/best-ai-gateways.md)
- [How to evaluate an AI gateway with a proof-based scorecard](https://keeprouter.com/blog/evaluate-ai-gateway.md)
- [Managed vs self-hosted AI gateways](https://keeprouter.com/compare/managed-vs-self-hosted-ai-gateways.md)
- [Is an AI gateway secure?](https://keeprouter.com/answers/is-an-ai-gateway-secure.md)
- [Does an AI gateway add latency?](https://keeprouter.com/answers/does-ai-gateway-add-latency.md)
- [models](https://keeprouter.com/models.md)

## Test the contract with a real model

Create a narrowly scoped key, select a model from the live catalog, and run the exact request shape your application depends on.

[Create a free key](https://keeprouter.com/login?returnTo=%2Fconsole%2Fkeys%3Fmodel%3Dfree) · [Live models and pricing](https://keeprouter.com/models.md)