LLM gateways,
compared honestly
Seven products that put models behind one endpoint, and one that does not. If inference is your only problem, pick from the first group. Squiid is for the other fifteen categories your app needs.
Short answer. Every product on this page solves some version of too many model providers. They differ on whether they resell inference or proxy your own accounts, whether they route between providers, whether you can self-host and what they charge. Squiid belongs in the list only for contrast: it does not route models, it provisions and bills the whole stack.
The one-line difference. A gateway answers which model should this request go to, and connects your agent to brains. Squiid answers what does this app need to exist: the database, hosting, email, SMS, storage, auth and payments, plus the brains.
The field at a glance
| Product | What it is | Who pays the provider | Self-host | Best for |
|---|---|---|---|---|
| OpenRouter | Reseller and router across 500+ models with cross-provider fallback. | They do, at par, with a platform fee at top-up. | No | Model breadth and routing at a low fee. |
| Inworld Realtime Router | One endpoint over 220+ models with CEL routing rules, A/B tests and TTS. | They do, with 0% markup on routed third-party models. | No | Per-user routing rules and experiments. |
| Vercel AI Gateway | The model layer of the Vercel platform, wired into the AI SDK. | Billed through your Vercel account, as of their public docs. | No | Apps already deployed on Vercel. |
| Cloudflare AI Gateway | A proxy in front of providers you already hold, with caching and analytics. | You do, directly. | Runs on Cloudflare | Caching, logging and rate limiting. |
| Portkey | An AI gateway with virtual keys, budgets, guardrails and observability. | You do, directly. | Yes | Governance over model traffic. |
| LiteLLM | Open-source SDK and proxy over 100+ model APIs in OpenAI format. | You do, directly. | Yes | Full control, no intermediary. |
| Helicone | LLM observability, as a proxy or an async logger. | You do, directly. | Yes | Seeing what the agent actually sent. |
| Squiid | Not a gateway in this sense. One account, key and bill for the whole stack across sixteen categories. | Squiid does, at par, with a fee at credit top-up. | No | Everything an app needs, not just models. |
Every comparison in this group
Squiid vs OpenRouter
ComparisonCheaper for pure inference and better at routing. Squiid covers the other fifteen categories.
Squiid vs Inworld Router
ComparisonBetter router, 0% on routed models, models only. Squiid covers the other fifteen categories.
Squiid vs Vercel AI Gateway
ComparisonPerfect if you live on Vercel and only need models. Squiid is host-neutral and covers the rest.
Squiid vs Cloudflare AI Gateway
ComparisonA proxy for accounts you already own, with caching and analytics. Squiid opens the accounts instead.
Squiid vs Portkey
ComparisonThe strongest model-layer controls of the group. Still models only, and you bring the accounts.
Squiid vs LiteLLM
ComparisonFree, self-hosted, excellent model budgets. You still open every account and run the server.
Squiid vs Helicone
ComparisonObservability, not procurement. Nothing in Squiid replaces a proper request log.
The three kinds of gateway
The first kind resells inference. OpenRouter, Inworld Realtime Router and the routers in that family hold the provider relationship, bill you for tokens and take a fee or a margin somewhere. You get one account and one catalogue, and routing between providers becomes a configuration choice rather than a code change.
The second kind proxies accounts you already hold. Cloudflare AI Gateway, Portkey, LiteLLM and Helicone sit in front of your own provider keys and add caching, budgets, guardrails, retries or observability. They do not reduce the number of accounts you have; they make the ones you have better behaved.
The third kind is platform bound. The Vercel AI Gateway is excellent if you are on Vercel and irrelevant if you are not, which is true of every host-attached service and worth saying plainly.
Where Squiid sits
Squiid is not competing in any of those three. It calls a named provider through a named path, with no cross-provider routing, no guardrails and no prompt store. What it adds is the layer underneath: it opens the accounts, holds the credentials, meters the calls against prepaid credits and sends one invoice, across databases, hosting, email, SMS, storage, auth, observability and payments as well as models.
If models are the only thing you buy, pick a gateway from the table above and do not buy Squiid. If your app needs a stack, a gateway solves one sixteenth of the problem.
The combination most teams end up with
One router or proxy for model traffic, chosen on catalogue and fee, plus Squiid for everything else. Two environment variables, two invoices at worst, and no provider credentials in the repository. That is a better answer than either product alone, and we would rather recommend it than pretend the categories overlap.
Questions people ask
Which LLM gateway is cheapest?
For pure resale, the lowest published fee in this group is 0% on routed third-party models at Inworld Realtime Router, with OpenRouter at 5.5% on its Standard tier as of 2026-09-18. Self-hosted LiteLLM has no fee at all if you ignore the cost of running it. Squiid is not the cheapest way to buy tokens and does not claim to be.
Is Squiid an LLM gateway?
Only in the narrow sense that model calls pass through its gateway. There is no cross-provider routing, no failover between vendors and no prompt management. Compare it on provisioning, key custody and billing across the whole stack instead.
Do I need a gateway at all?
If you call one model from one provider, no. A gateway earns its place when you want to switch models without a deploy, fail over when a provider degrades, cap what each key can spend, or see what your agents actually sent.
Can I run a gateway and Squiid together?
Yes, and it is the setup we recommend to anyone who cares about routing. Model traffic to the gateway, everything else to Squiid. See the gateway convention for how the Squiid side looks.
Which one should a solo vibe coder use?
Start with whatever your agent already supports, usually a single provider key. Add a router when model choice starts mattering, and add Squiid when the number of accounts, not the number of models, becomes the annoyance.
Pick a gateway for brains.
Pick Squiid for the rest.
One account, one key and one bill for the whole stack. No markup on provider prices.