Guides

Written guides

The catalogue answers what each product does. These guides explain how to choose one, what it costs, and how to put its controls to work in production.

All guides

  1. 01 I keep seeing the term. What is it, and what is it not? What is an LLM gateway? A plain definition of an LLM gateway, and how it differs from a router, a proxy, a traditional API gateway and an MCP gateway — five terms vendors use interchangeably and should not.
  2. 02 Just tell me which one to use. Best LLM gateways in 2026 There is no single best LLM gateway. Five common jobs — self-hosting, fastest start, HIPAA, the Vercel stack, one hyperscaler — and the product we would pick for each, with the deciding figures linked.
  3. 03 What will this actually cost me on top of the tokens? How LLM gateway pricing works Token markup, credit top-up fees, per-seat pricing and BYOK are four unrelated charging models sold under one word. What each costs, and which one wins at your spend.
  4. 04 Should I run this myself or let someone else? Self-hosted vs managed LLM gateways ‘Self-hostable’ describes where the binary runs, not what it may do. Which governance features sit behind a paid edition, what air-gapped deployment really requires, and what running one costs beyond the licence.
  5. 05 If the guardrail breaks, does my request still go through? How LLM gateway guardrails fail A guardrail is a security control, and most products here do not publish what happens when the checker itself fails. Where the checks run, what they block by default, and what to ask before you depend on one.
  6. 06 What happens when a provider goes down? How LLM gateway failover actually works Understand LLM gateway retries, timeouts and fallback targets. Build a request deadline, test streaming failures, and avoid duplicate tool actions.
  7. 07 Can I put this in front of a security review? LLM gateway compliance: SOC 2, HIPAA and evidence Evaluate LLM gateway SOC 2 evidence, HIPAA BAAs, retention and subprocessors. Separate a vendor claim from an audit report and check the service in scope.
  8. 08 How hard would it be to leave later? How much does an LLM gateway lock you in? OpenAI-compatible does not mean interchangeable. Compare API contracts, model IDs, policies and data exports, then test a migration before committing to an LLM gateway.
  9. 09 Will this vendor still be here next year? Who still owns your LLM gateway? Products in this catalogue changed hands during 2026, and many are proprietary with no repository to fork. What an acquisition or a quiet abandonment does to an integration in your request path.
  10. 10 What happens if it goes down and it is not my fault? Does your LLM gateway promise any uptime? Only a minority of these products publish an uptime figure, and for a self-hosted gateway its absence is structural rather than evasive. What each published number permits in downtime per month.
  11. 11 Who can read what I send? Which LLM gateways store your prompts? Some of these products store full prompt and completion text by default. Opt-out exists almost everywhere, but the default is what ships, and ‘not stored’ is not the same as ‘not sent’.
  12. 12 Why does every vendor claim a different number? LLM gateway model counts: what “500+” means Model and provider counts are not comparable between vendors, and several vendors publish more than one number for themselves. Why this catalogue records a range, and what to count instead.
  13. 13 Will this show up in the tools I already run? LLM gateway observability: traces, logs and export Compare LLM gateway tracing and export support. Check plan restrictions, span contents, privacy and application context before connecting an OpenTelemetry collector.
  14. 14 Will the gateway actually stop my agent when the money runs out? LLM gateway spending limits: stop a runaway agent bill? Compare LLM gateway spending limits: what they block, whether BYOK counts, how concurrent requests overshoot, and how to test the cutoff before an agent runs unattended.
  15. 15 Is a cheaper route making repeated prompts more expensive? Is routing destroying your prompt cache? Learn when LLM routing reduces prompt cache hits, how sticky routing helps, and how to compare cached-token costs with a repeatable routing test.
  16. 16 How can a team change the model behind an LLM gateway without breaking production workflows? Changing models without breaking production Plan a model migration with a version inventory, contract tests, quality gates, canary traffic and an executable rollback. Distinguish aliases from real model identity.
  17. 17 What must work before a team can safely route coding agents through an LLM gateway? Running coding agents through an LLM gateway Check coding-agent gateway compatibility beyond a base URL: API formats, streaming, tools, model aliases, context, credentials and team rollout controls.
  18. 18 Do agents need an MCP gateway in addition to an LLM gateway? Do you need an MCP gateway as well? Decide whether your agents need an MCP gateway. Compare tool discovery, authorization, upstream credentials and audit trails with a practical acceptance checklist.
  19. 19 What should a team evaluate when choosing an LLM gateway for realtime voice agents? Choosing a gateway for realtime voice agents Evaluate live voice gateways for session authentication, interruptions, tool calls, reconnection and cost. Includes a conversational latency and failure-test worksheet.
  20. 20 How do I find every application affected by a model alias change or API shutdown? Model aliases and deprecations: build an endpoint inventory Track aliases, snapshots, SDK dependencies and shutdown dates so an upstream model retirement does not surprise production.
  21. 21 How should I choose a hosting provider for a Chinese-developed model? Choosing a provider for Qwen, DeepSeek, GLM or Kimi Compare Chinese-developed model endpoints by exact version, serving configuration, price, privacy and task performance.

Or skip the reading

The guides are opinions with their evidence attached. If you would rather work from the evidence directly, three tools do the same job without the prose.

Find your fit

Six questions about your constraints, then a filtered shortlist that shows why each product was ruled in or out.

Compare side by side

Any two products, every field, differences highlighted — plus written verdicts on the pairs people actually weigh up.

Estimate the cost

Your volumes and team size against all 31 products, broken into named lines with assumptions flagged.

How these are written

Each guide carries a named byline and a published and updated date read from that page's own revision history, not from the last site deploy. Editorial recommendations are distinguished from vendor claims. Catalogue figures render at build time with their source links and check dates attached. Detailed implementation reviews cite vendor documentation with a separate research date, and illustrative calculations state their assumptions.

Nobody pays to be included, recommended or ranked, and there are no affiliate links. The methodology sets out what is scored and what this site deliberately refuses to score; about says who writes it.