If you've built anything that calls an AI model, you've probably hit the same wall: your code is wired directly to one provider's API. Then that provider has an outage, or raises prices, or a competitor ships a model that's suddenly better and cheaper for your use case — and switching means rewriting the part of your app that talks to the model. An AI gateway (also called an LLM router) sits between your app and every AI provider behind a single, stable API. You keep writing code the same way; the gateway decides, request by request, which model actually handles it — and can automatically fall back to a different one if your first choice is down, slow, or too expensive. Here are six real gateways, from the free self-hosted option to the enterprise-ready ones, with honest pricing and trade-offs.
1. Helicone — the gateway that comes with observability built in
Helicone is the pick if you want more than routing: it's an open-source AI gateway and an LLM observability platform in one, so every request you route also gets logged with cost and latency data automatically. Integration is a one-line proxy change, and it covers 100+ models through a single API.
Pricing: free for 10,000 requests/month, usage-based paid plans beyond that, and the self-hosted version is free under Apache 2.0.
Good for: teams who want to see exactly what their AI features cost and how fast they respond, not just route traffic blindly.
Watch out for: the proxy becomes a dependency in your request path, and self-hosting means running five separate services, not one container.
2. Requesty — the broadest model catalog with enterprise controls
Requesty gives you one OpenAI-compatible endpoint for 600+ models across 30+ providers — the widest catalog of the tools here. Pricing is a flat 5% markup on whatever the underlying model costs, which is easier to reason about than some competitors' tiered plans, and it includes $10 of free credit to test with.
Pricing: free tier at 200 requests/day (free models only); pay-as-you-go beyond that at a 5% markup, no subscription required; enterprise plans add SSO, RBAC, audit logs and EU data residency.
Good for: teams that want enterprise governance (SSO, audit logs) without giving up pay-as-you-go pricing.
Watch out for: it's a young company competing against already-entrenched players like OpenRouter and Portkey, and its usage stats are self-reported.
3. OmniRoute — free, open-source, self-hosted
OmniRoute is a traffic-control layer you run yourself: fully free, MIT-licensed, and built to automatically switch to a backup provider if one goes down, while also trimming your prompts to cut token usage (it claims up to 95% compression in some cases). It plugs into 33+ coding tools including Claude Code, Cursor and Copilot, with an MCP server built in.
Pricing: completely free and open source, self-hosted.
Good for: teams comfortable running their own infrastructure who don't want to pay a per-request commission to a third-party gateway.
Watch out for: no SLA, community support only — a real risk if this sits in front of production traffic that actually matters. It's also a very new project (started early 2026), so there's little long-term track record yet.
4. Millwright — for teams that don't want their API keys leaving their infrastructure
Millwright takes self-hosting further than OmniRoute: it's a small, free, open-source tool you run entirely on your own infrastructure, so your provider API keys and traffic never touch a third party. It automatically picks the cheapest model that can do the job and fails over with circuit breakers if a provider goes down, plus it tracks spend in a local SQLite or Postgres database.
Pricing: free, Apache-2.0 licensed. You only pay for your own hosting and your provider API keys.
Good for: security-conscious teams who specifically don't want a hosted gateway seeing their traffic.
Watch out for: this is a very young project — under 30 GitHub stars and about 5 commits at last check. It's far less battle-tested than an established router like LiteLLM, and you're on your own for support.
5. Opper AI — built for GDPR compliance from day one
Opper AI is a Europe-based gateway (Sweden/EU) giving access to 300+ models through one OpenAI-compatible API, with data kept on EU servers and AES-256 encryption at rest — including an optional Zero Data Retention mode for enterprise customers. Failed or fallback requests aren't billed.
Pricing: pay-as-you-go — 3% fee on the gateway alone, or 5.5% plus usage-based billing if you also want their observability/routing control plane; enterprise is custom with volume discounts.
Good for: teams based in Europe, or serving European customers, who need a documented GDPR story more than they need the absolute cheapest routing.
Watch out for: percentage-based fees can get expensive at high volume, and outside a European compliance context there's less reason to pick it over the alternatives here.
6. Auriko — routes purely on price
Auriko's pitch is narrower and more aggressive than the others: it checks prices across a dozen-plus providers in real time and sends each request to whichever is cheapest right now, without you changing any code, plus automatic failover and workspace-level budget controls. It claims roughly 30% average savings versus calling one provider directly.
Pricing: not published on the site — you have to sign up to see actual rates, which is a friction point compared to the others on this list.
Good for: teams whose main pain point is the inference bill, not observability or compliance.
Watch out for: unpublished pricing means you can't compare costs upfront, the 30% figure is a marketing claim that will vary with your real traffic, and there's no public information on funding or the team behind it.
How they compare
| Tool | Models | Pricing model | Best for |
|---|---|---|---|
| Helicone | 100+ | Free up to 10k req/mo, then usage-based | Gateway + observability combined |
| Requesty | 600+ | 5% markup, pay-as-you-go | Broadest catalog + enterprise governance |
| OmniRoute | Dozens | Free, self-hosted (MIT) | Zero-cost, DIY infrastructure |
| Millwright | Provider-dependent | Free, self-hosted (Apache-2.0) | Keeping keys and traffic in-house |
| Opper AI | 300+ | 3-5.5% fee + usage | GDPR / EU data residency |
| Auriko | 12+ providers | Undisclosed, sign-up required | Squeezing the lowest per-call price |
If you just want the safety net of failover without ripping out your stack, Helicone or Requesty are the easiest hosted starting points. If you'd rather not hand your traffic to a third party at all, OmniRoute and Millwright do the same job self-hosted for free — just budget time for the operational overhead that comes with running it yourself. And if the whole reason you're looking is a scary inference bill, Auriko and Requesty's transparent-markup model are worth testing against your actual usage before you commit to anything.