Hosted LiteLLM alternative: the gateway without a proxy to run

A hosted LiteLLM alternative gives you the gateway without running the proxy. You keep one base URL, a key per tool, budgets and fallbacks, and you drop the server you patch. On 2026-10-21 LiteLLM's release pipeline published a malicious 1.107.0 that sat on PyPI from 11:39 to 13:23 UTC (its incident post, read 2026-10-22). It follows a PyPI compromise in March and a proxy CVE in April.

This page dates those three posts, maps a LiteLLM setup to a hosted key feature by feature, and prices one month at list through LiteLLM and through the live catalog here. For the whole field, see every gateway ranked by the fee it adds. On 2026-10-22, 58 of the 75 chat models with a list price sit under it here on input and output.

LiteLLM is a proxy you run yourself. Its pricing page puts the open-source gateway at $0, free forever and self-hosted, and says your data and your keys never leave your own infrastructure (read 2026-10-22). That is its strength. It is also the job: you deploy the proxy, upgrade it, back up its database and hold every provider key it calls.

A hosted gateway moves that job off your servers. Four things change.

• Who holds the provider keys. Through LiteLLM, your Anthropic key, your OpenAI key and the rest sit in its config or its database. Here you hold one sk-uplink- key, and UplinkSwitch holds what sits behind it.
• Who patches the gateway. A self-hosted proxy is yours to pin, upgrade and rebuild when an advisory lands. A hosted one has no gateway package on your hosts to poison.
• What a token costs. Through LiteLLM each provider bills your own key at its list price. Here the live catalog prices 58 of the 75 chat models with a list price under it.
• What you can check. The daily real-model check sends every model the same prompt and publishes the verdict (236 of 273 stable at the last run). Every non-streamed answer carries a signed receipt that the public verifier checks in the browser.

Your code does not change. LiteLLM speaks the OpenAI format and so does the hosted key, so the same SDK call works once two environment variables point somewhere else.

LiteLLM published each incident with the versions, the time window and the fix, and updated the March and October posts as it learned more. The table quotes those posts as read on 2026-10-22. Check them for updates: the October investigation was still open when we read it, and its post was last updated at 20:30 UTC on 2026-10-21.

The October post asks anyone who ran 1.107.0 to rotate every provider key stored in the LiteLLM proxy database, plus the master and virtual keys and the database password. That list is what holding provider keys on a host costs after an incident: one bad release puts every key in scope at once.

A hosted key narrows the list to one. If a laptop or a CI runner leaks an sk-uplink- key, you revoke it in the console, and the revocation takes effect on the next request. The provider keys behind it never sat on your machines.

Pinning helps on either side. LiteLLM's posts recommend an explicit version, the GitHub advisory names the fixed release for the April issue, and the litellm page on PyPI lists the releases you can pin to.

Six hosted gateways, read on the dates in the table. Two columns matter most here. The first says who holds the provider keys, which decides what you rotate after an incident. The second is the fee on top of the developer's list price.

OpenRouter and Requesty hold the provider keys by default and add 5.5% and 5% to list. Portkey and Cloudflare sit in front of your own keys, so those keys stay yours to rotate. Vercel sells at list. UplinkSwitch holds the keys and prices under list. Portkey has been part of Palo Alto Networks since the deal closed on 2026-05-29; its plan math is on the Portkey alternatives page, and Requesty's 5% on a coding month is on the Requesty alternatives page.

The left column is what LiteLLM's pricing page lists for the free, self-hosted gateway (read 2026-10-22). The right column is the same job on a hosted UplinkSwitch key, from the console and the API.

Each hosted key can also carry a model allowlist and an expiry date. Narrow a CI key to two models, give it a week, and a request for any other model returns 403 scope_denied. What each request stores, and what it never stores, is in the no-logs post.

A self-hosted gateway bills in two parts. Tokens cost each provider's list price, and the server, its database and your time cost whatever they cost you. A hosted gateway folds the second part into its price. The tables below price the tokens and leave your server as the one number only you know.

The mix is four Anthropic and OpenAI models, each at 1,000,000 input and 200,000 output tokens. On these the hosted resellers bill the developer's own list price, so list plus each fee rule is what each one charges:

On this mix, $500 of list-price usage costs $211.11 here and $5,000 costs $2,111.11, with no server to add. Through LiteLLM the tokens cost the full $500 or $5,000, then the server. OpenRouter and Requesty add their fee to the same list price. Per model the gap varies: Claude Sonnet 5.5 is $1.20 in and $3.00 out here against Anthropic's $2.00 and $10.00.

Your mix will differ. The LLM API price comparison puts every chat model's list price next to ours, with a calculator for your own token counts.

• Data must stay inside your network. LiteLLM runs in your cloud, so prompts and responses never leave your environment (its Enterprise docs, read 2026-10-22).
• You hold negotiated provider contracts or committed spend. LiteLLM routes through your own keys at your own rates.
• You need a provider no hosted gateway lists. LiteLLM covers 100+ providers behind one OpenAI API, self-hosted models included.
• You want the code. The gateway is open source and free forever, with no license fee in production (its pricing FAQ, read 2026-10-22).
• You need SSO and SCIM, audit logs, org and team admins, or an air-gapped install. Those come with LiteLLM Enterprise, priced by annual request capacity and never per token, with a 30-day trial key.

• You would rather not run the proxy. There is no container to pin or upgrade, and no database of provider keys on your hosts.
• You want prices under list: 58 of the 75 chat models with a list price sit under it on input and output today, Claude Sonnet 5.5 at $1.20 in and $3.00 out against $2.00 and $10.00.
• You want proof of the model. The daily check at /proof sends every model the same prompt and publishes the verdict (236 of 273 stable at the last run).
• You want records you can check without trusting the gateway. Every non-streamed answer carries a signed receipt, monthly statements are signed too, and statements/verify checks either one in the browser.
• You want spend control per key without a config file: a monthly limit, budget autopilot near the limit and rate limits per key, all set in the console.
• You pay as you go. Credit is prepaid from $10, with no subscription and no fee on top-ups, and each top-up locks your prices for 30 days.
• You run Claude Code or the Anthropic SDK. The Claude models also answer in the Anthropic format, at the same price as the OpenAI format; the SDK setup guide covers all three formats.

There are two ways. To drop the proxy, point the OpenAI SDK at the hosted base URL and swap the key. Model ids are on the Models page.

# Before: the app called your LiteLLM proxy
export OPENAI_BASE_URL="http://0.0.0.0:4000"
export OPENAI_API_KEY="sk-<your-litellm-api-key>"

# After: the hosted key, no proxy
export OPENAI_BASE_URL="https://api.uplinkswitch.com/v1"
export OPENAI_API_KEY="sk-uplink-live-..."

To keep LiteLLM for its routing and budgets, add the hosted key as one of its providers. LiteLLM's docs for OpenAI-compatible endpoints use the openai/ model prefix, an api_base and an api_key, with the /v1 suffix on the api_base (read 2026-10-22).

model_list:
  - model_name: claude-sonnet-5-5
    litellm_params:
      model: openai/claude-sonnet-5-5
      api_base: https://api.uplinkswitch.com/v1
      api_key: sk-uplink-live-...
  - model_name: glm-5.3
    litellm_params:
      model: openai/glm-5.3
      api_base: https://api.uplinkswitch.com/v1
      api_key: sk-uplink-live-...

Behind LiteLLM, keep the version pinned. One hosted key in its database replaces several provider keys, so after an incident there is one key to revoke instead of a list. The SDK guide has the Anthropic and Gemini formats for tools that need them.

• Who holds the provider keys, and what you rotate if a key leaks. A gateway that runs over your own keys leaves them on your list.
• The fee on top of list, and whether it applies to credit purchases or to each request. The OpenRouter vs LiteLLM cost math works the break-even at three monthly spends.
• Whether the model served can be checked. Our verdict per model is public at /proof, and GET https://api.uplinkswitch.com/v1/proof gives the canary prompt so you can replay it against the developer's own API.
• Route health. The status page shows the API and its routes from hourly probes over the last 30 days, with the same data as public JSON.
• Per-key limits. A monthly spend limit stops a runaway agent at a number you set, and a model allowlist keeps a CI key on the models it needs.

What is the best alternative to LiteLLM?

It depends on why you are leaving. To keep the gateway inside your own network, the self-hosted open-source gateways stay closest to LiteLLM. To stop running a proxy, pick a hosted gateway: OpenRouter adds 5.5% to list on credit purchases and Requesty adds 5% to model cost (both read 2026-10-22), Portkey and Cloudflare AI Gateway sit in front of your own provider keys, and UplinkSwitch prices 58 of the 75 chat models with a list price under it, with no fee on top-ups.

Is LiteLLM free?

Yes. The open-source gateway is $0 and free forever to self-host, with no license fee in production; each provider bills you at its list price, and you run the server. LiteLLM Enterprise adds SSO and SCIM, audit logs and 24/7 support, priced by annual request capacity and never per token, with a 30-day trial key (litellm.ai/pricing, read 2026-10-22).

Is there a hosted version of LiteLLM?

LiteLLM's pricing page listed two plans on 2026-10-22, Open Source and Enterprise, and both are self-hosted; its Enterprise docs say it runs in your cloud. The usual swap is a hosted gateway that speaks the same OpenAI format: change the base URL and the key, and the SDK code stays as it is.

Was LiteLLM compromised in 2026, and what should I do?

Three of LiteLLM's 2026 posts describe security incidents: litellm 1.82.7 and 1.82.8 on PyPI on 2026-03-24, CVE-2026-42208 in the proxy's key check (published 2026-04-29, fixed in 1.83.7), and a malicious 1.107.0 published by its release pipeline on 2026-10-21. For the October release LiteLLM advises pinning 1.106.2, rebuilding, and rotating every provider key stored in the proxy database, the master and virtual keys and the database password (its post, read 2026-10-22).

Which LLM gateway is the best?

The one that matches where your keys and your data have to live. Self-host LiteLLM when traffic must stay inside your network or you hold negotiated provider contracts. Use a hosted gateway when you want no server at all: OpenRouter for the largest model list, Requesty for EU data residency and routing policies, UplinkSwitch for most chat models under list with a daily real-model check and signed receipts.

Can I keep LiteLLM and use a hosted key behind it?

Yes. LiteLLM routes to any OpenAI-compatible endpoint with the openai/ model prefix, an api_base that ends in /v1 and an api_key (its docs, read 2026-10-22). Point it at https://api.uplinkswitch.com/v1 with an sk-uplink- key, and LiteLLM keeps its virtual keys and budgets while one hosted key replaces the provider keys in its database.

LiteLLM, OpenRouter, Requesty, Portkey, Palo Alto Networks, Cloudflare and Vercel are trademarks of their owners. This page quotes LiteLLM's pricing page, Enterprise docs, provider docs and three security posts, the GitHub advisory, and the OpenRouter, Requesty and Portkey pricing pages, all read on 2026-10-22, and the Cloudflare and Vercel pricing docs as read on 2026-10-19. Our figures come from GET https://api.uplinkswitch.com/v1/catalog, GET https://api.uplinkswitch.com/v1/proof and the terms. Get a key: prepaid from $10, no subscription, 30-day price lock, in the console.