All three put one OpenAI-compatible API in front of many model providers, run your requests on your own provider keys, and fall back when a provider fails. They differ in who runs the gateway, what you pay for, and what comes on which plan. Portkey’s and LiteLLM’s sides are taken from their own public pages, each linked, as they read on 2026-10-24.
Side by side
| Autonomous Relay | Portkey | LiteLLM | |
|---|---|---|---|
| What it is | A hosted gateway. Self-hosting is an Enterprise option, not open source. | A hosted gateway, now Palo Alto Networks’ Prisma AIRS AI Gateway, and an open-source gateway you can host yourself. source | An open-source (MIT) proxy you run yourself. source |
| Price | Solo free; Pro $29, Team $99 and Business $399 a month per organisation; Enterprise custom. | Developer free; Production $49 a month; Enterprise custom. source | Free to self-host, on your own servers and Postgres; an Enterprise licence priced annually, not per token. source |
| What usage costs | Nothing per request or log; priced per organisation. | 10,000 recorded logs a month on Developer, 100,000 on Production, then $9 per extra 100,000. source | The infrastructure you run it on. source |
| Tokens | On your own provider keys, at 0% markup. | On your own provider keys. source | On your own provider keys. source |
| Fallbacks | A models chain (the first three models, three attempts), plus account-wide defaults. | A fallback config, on any non-2xx answer by default or the status codes you choose; every plan. source | Per model group after retries; also context-window and content-policy fallbacks. source |
| Spend limits | Per key, agent and workspace on every plan; guardrail budgets on Team. | Budgets on keys; granular budgets and rate limits on Enterprise. source | Per key, user and team, free (they need Postgres); per-model budgets on Enterprise. source |
| PII redaction and prompt-injection detection | Built in and deterministic, on every plan; custom patterns and block mode on Team. | Basic guardrails on Developer, partner and pro guardrails on Production; the prompt-injection scan is LLM-based. source | LLM guardrails free; some integrations, Lakera prompt injection among them, need Enterprise. source |
| Single sign-on and SCIM | Enterprise. | Enterprise. source | With an Enterprise licence. source |
When Portkey is the better fit
- You want an open-source gateway you can also run yourself, with a hosted service beside it.
- You want guardrails from partner platforms alongside the built-in checks.
When LiteLLM is the better fit
- You want to run the gateway on your own infrastructure, under an open-source licence.
- You want no vendor in the request path at all, and can operate the proxy and its database.
When Autonomous Relay is the better fit
- You want nothing to run: a hosted gateway, with self-hosting only if an Enterprise agreement calls for it.
- You want a flat price per organisation that does not grow with your request or log volume.
- You want deterministic PII redaction, prompt-injection detection and spend limits on every plan, the free one included.
Switching
All three speak the OpenAI Chat Completions format, so moving is a base URL and a key: point the OpenAI SDK at https://autonomousrelay.com/v1, add your provider keys, and keep your code. See the quickstart, fallbacks and failover, spend limits and pricing, or how it compares with OpenRouter.