TierUp vs LiteLLM: run the gateway yourself, or stop choosing models

Here's something comparison pages rarely admit: these two products barely compete.

LiteLLM is an open-source proxy and Python SDK — 52k+ GitHub stars, 100+ providers, used in production at companies like Netflix (their site's claim, not ours). You deploy it, you point it at your provider keys, and you get one OpenAI-shaped interface to every model you care about, plus virtual keys, budgets, load balancing, fallbacks, and guardrails. It's genuinely good infrastructure. It also assumes something about you: that you want to choose models, and you're willing to operate a service to do it well.

TierUp assumes the opposite. You send requests to tier-1 through tier-4 (Speed, Balance, Intelligence, Reasoning) and we route each one to what we assess as the best-value model in that class right now. No model names — responses strip them out so your code can't grow a dependency on a checkpoint. No proxy to deploy. Also, candidly, no self-hosting option, no SSO, no admin dashboard for your team, and ~zero production users. We're a hosted abstraction in public beta, built on top of OpenRouter, and we say so on every page including this one.

So the real question isn't "which gateway is better." It's: do you want to own the model decision (and the box it runs on), or delegate both?

Side by side

TierUpLiteLLM
What it isHosted API with 4 performance tiers; we pick the model behind each tierOpen-source proxy server + Python SDK giving one interface to 100+ providers; you pick the models
Who picks the modelWe do — tier mappings are versioned server-side, and responses strip provider/model detailsYou do. LiteLLM routes to whatever you configure, with fallbacks and load balancing you define
Where it runsOur infrastructure (which routes via OpenRouter today)Yours — your cluster, your Redis, your Postgres, your pager
Ops burdenNone on your side beyond an API keyReal: deploy, upgrade, monitor, and scale the proxy yourself, or pay for their enterprise/cloud offering
Model churnOurs. A better model ships, we re-map the tier, your code doesn't changeYours. LiteLLM makes swapping models easy, but deciding when and what to swap to is still your job
Software costNo platform fee; you pay per tokenFree and open source; enterprise tier (SSO, audit logs, support) is custom-priced
Token costFlat per-tier, ~50% under the retail price of the underlying models (tier 2: $1.50/$7.50 per 1M in/out vs ~$3/$15 retail Sonnet-class). Transparently subsidized while we test product–market fit. Tier 1 currently freeRetail — you bring your own provider keys and pay each provider's list price
Spend controlsPrepaid wallet, per-request usage logging, daily spend guardrailVirtual keys, per-key/team/org budgets, rate limits, cost tracking — considerably richer than ours
SDK compatibilityOpenAI-compatible: change the base URL, set model to a tierOpenAI-compatible by design; also callable as a Python SDK
MaturityPublic beta, built solo, ~zero users. Not a typo52k+ stars, huge community, serious production adoption. On maturity, LiteLLM wins without breaking a sweat

When to use which

Use LiteLLM if:

  • You have platform or infra engineers and you want the gateway inside your own network, under your own control, on your own keys.
  • You need per-team budgets, virtual keys, SSO, audit logs — org machinery TierUp simply doesn't have.
  • Model choice is part of your product's edge. If you benchmark and pin checkpoints, a tier abstraction would just be in your way.
  • You're already at the scale where running your own gateway pays for itself. Plenty of teams are.

Use TierUp if:

  • You'd rather not stand up Redis, Postgres, and a proxy deployment to call a language model.
  • Your actual requirement is a performance class — "fast and cheap," "balanced," "smartest available" — not a specific checkpoint.
  • You want model churn to be someone else's job. When a better model ships, your tier upgrades server-side; there's no config to edit because there's no config.
  • You want to test that idea for free: tier 1 costs nothing right now, and signup comes with $25 credit, no card.

Use both, plausibly: LiteLLM proxies any OpenAI-compatible endpoint, and TierUp is one. Some teams will route most traffic through models they chose, and point one LiteLLM deployment at a tier for workloads where the class matters more than the name. We'd count that as a win, not a compromise.

The economics, stated plainly

LiteLLM's software is free; your costs are provider tokens at retail plus the infrastructure and engineering hours to run the proxy well. TierUp charges no platform fee and prices tiers ~50% below the retail cost of the models behind them — and we currently eat that difference, on purpose, capped by a daily spend guardrail, while we find out whether developers will hand model choice to a router. That subsidy is a funded experiment, not a business model. If we earn long-term margin, it comes from volume pricing and routing optimization. You should know that before you build on us, so: now you do.

FAQ

Can I self-host TierUp the way I'd self-host LiteLLM?

No. TierUp is hosted-only, and the tier abstraction is the product — the server-side mapping of tiers to models is the part we operate for you. If your security posture requires the gateway inside your own network, LiteLLM is the right call and we won't argue.

Isn't LiteLLM free? How is TierUp cheaper?

The software is free; the tokens aren't. With LiteLLM you pay every provider's retail price plus whatever it costs you to run the proxy. TierUp's tiers are priced ~50% under retail because we subsidize them during product–market-fit testing — transparently, with a cap. Cheaper today, yes; permanent, we can't promise.

Does TierUp have LiteLLM's team features — virtual keys, per-team budgets, SSO?

Mostly no. We have a prepaid wallet, per-request usage logging, and a daily spend guardrail. LiteLLM's access-control and budgeting machinery is far deeper. If you're provisioning an organization rather than an app, that gap probably decides it.

Which model does my tier request actually hit?

We don't say, deliberately — responses strip model and provider details so nothing in your stack can depend on a checkpoint. We commit to the class instead: tier 1 speed-class, tier 2 balance-class, tier 3 intelligence-class, tier 4 reasoning-class. If you need to know and pin the exact model, use LiteLLM (or a direct API); that need is exactly what TierUp is built to remove, so it's a poor fit for you by design.

Also see: TierUp vs OpenRouter · LLM API cost calculator

Try a tier before you believe any of this.

The playground at tierup.ai/try works without signing up — no proxy to deploy, no keys to bring. If you like it, signup gets you $25 in credit, no card required.

Featured on Fazier