ngrok AI Gateway Review (2026): Pricing, Features & Honest Verdict

ngrok AI Gateway is an LLM routing layer that puts every model you call behind one private endpoint. It is a clean piece of infrastructure with unusually honest metering, held back by a thin built-in provider list. Best for: backend teams juggling several model providers. Price: from $0 (free ngrok plan), plus $0.05 per million tokens. Free plan: yes. Rating: 7.5/10.

What is ngrok AI Gateway?

ngrok AI Gateway launched on 5 August 2026 and landed at number four on Product Hunt that day. It gives your application a single endpoint, gateway.ngrok.ai, and routes whatever you send there to OpenAI, Anthropic, a self-hosted model, or a custom provider. Your code stops knowing or caring which one answers.

The company behind it is not a newcomer. ngrok has spent years as the default way developers expose a local port to the internet, and the ingress product sits in a very large number of development workflows. This is that same team applying its networking instincts to model traffic. We found the design reflects that history: the gateway is treated as a network hop with credentials, policy and failover, not as an AI product with a chat window bolted on.

What are ngrok AI Gateway’s key features?

One endpoint for hosted and local models

The core promise is that your application calls one address regardless of what is actually serving the request. Cloud providers and self-hosted models sit behind the same door. You can route to Ollama or vLLM running on your own hardware alongside OpenAI and Anthropic, which matters if you are moving cheap high-volume work onto local models while keeping the hard requests on a frontier model.

Credentials that never reach your client code

This is the part we liked most. The gateway splits credentials in two. Your application authenticates with an ngrok access key, prefixed ng-, and the upstream provider key stays on the server side where the client never sees it. You can issue a separate access key per application or per team and revoke one without touching the others. Invalid keys get an immediate 401 Unauthorized. Tokens display once at creation, so a lost key means deleting and recreating rather than looking it up.

Anyone who has watched an OpenAI key leak into a mobile bundle will understand why this arrangement is worth something on its own.

Automatic failover between providers

If a configured provider fails, the gateway attempts the next one in line. This is basic reliability engineering, and it is the sort of thing every team eventually writes badly by hand. Having it in the routing layer, where retries and timeouts belong, is the right place for it.

No provider accounts required to start

You can call OpenAI and Anthropic through ngrok.ai inference without holding accounts with either. Credits cover both the gateway fee and the upstream model cost. For a prototype, or for a team that does not want three vendor relationships and three invoices before it has shipped anything, that removes real friction.

Drop-in SDK compatibility

Existing Python and TypeScript SDKs work by pointing the base URL at the gateway. There is no bespoke client library to adopt and no rewrite of your call sites. Migration is a configuration change, and just as importantly, migration away is also a configuration change. Gateways that are easy to leave are gateways worth trying.

How much does ngrok AI Gateway cost?

The gateway itself charges USD $0.05 per million tokens on every successful request, counted across combined input and response tokens. That is the processing fee, and it is the only figure ngrok adds when you bring your own provider keys. Credits require a minimum top-up of USD $5.00, and purchased credits expire one year after purchase. When your balance hits zero the gateway rejects new requests until you top up.

Sitting underneath that is the ordinary ngrok subscription. The Free plan is $0 and includes $5 of one-time usage, up to 3 online endpoints, 1GB of data transfer, 20,000 HTTP/S requests monthly and 1 team member. Hobbyist is $8/month billed annually or $10/month billed monthly, with $10 of monthly included usage, 5GB of transfer and 100,000 requests. Pay-as-you-go is $20/month with $20 of included usage, unlimited endpoints, 5GB of transfer then $0.10/GB, 100,000 requests then $1 per 100,000, and unlimited team members with 3 included. Enterprise is custom and covers SSO/SAML, SCIM, SLAs and security reviews.

Five cents per million tokens is a small number against what the models themselves cost. Compared with gateway competitors that take a percentage of spend, a flat per-token fee is more predictable and gets relatively cheaper as your volume grows. Self-hosted alternatives are free in licence terms and cost you an on-call rotation instead.

What is actually opaque about the pricing?

One thing deserves flagging because it is easy to miss. When you use your own provider keys, the maths is clear: ngrok takes $0.05 per million tokens and your provider bills you directly for the model. When you use ngrok.ai built-in inference, credits cover the processing fee and the model cost as a bundled rate, and the documentation does not itemise what portion is which.

So the convenient path, the one where you skip provider accounts entirely, is also the path where you cannot see your model costs. For a prototype that is fine. For anything running at scale it is not, and you will want to move to your own keys precisely so you can audit the bill. Treat built-in inference as a fast start rather than a destination, and be aware that the switch is one you should plan for rather than discover.

What is missing at launch?

A few things, and they are the sort of gaps you would expect from a gateway shipped in its first week rather than design mistakes.

The published documentation covers authentication, credits, provider routing and failover well. What it does not yet cover in detail is the observability layer most teams want from a gateway: per-model latency breakdowns, prompt and response logging with redaction controls, and cost attribution by application or team. The access key model gives you the raw material for that last one, since you can issue a key per application, but you are inferring cost from key usage rather than reading a report.

Semantic caching is the other notable absence. Several competing gateways will return a stored response for a repeated or near-identical prompt, which cuts both latency and spend on workloads with repetitive queries. If your traffic has that shape, the savings can dwarf the routing fee, and you will not find it here yet.

Rate limiting per key is the third. Being able to cap a runaway client before it burns your credit balance is the kind of control that becomes urgent the first time a retry loop ships to production. None of this makes the gateway a bad choice today, but it does mean that if your reason for adopting a gateway is governance and cost control rather than routing and credential hygiene, you should check the current documentation rather than assuming these arrived after launch.

Who is ngrok AI Gateway best for?

Use ngrok AI Gateway if you are calling more than one model provider from a backend and you have started writing your own routing, retry and key-management code. That code is a liability and this replaces it. It is a particularly good fit if you already pay ngrok for ingress, because the billing, team management and access model are things you have already set up.

It also suits teams running a hybrid setup, with a local model handling bulk work and a hosted frontier model handling the hard cases. Routing both through one endpoint keeps that split out of your application logic.

Skip ngrok AI Gateway if you only call one provider. A gateway in front of a single upstream adds a hop, a dependency and a fee in exchange for very little. Skip it too if you need broad provider coverage today: the built-in inference path documents OpenAI and Anthropic, and anything else means bringing your own keys. And if your organisation requires itemised per-model cost reporting from day one, the bundled inference rate will not satisfy your finance team.

What are the best ngrok AI Gateway alternatives?

OpenRouter is the obvious comparison and covers a far wider catalogue of models from many more providers through one API. It charges a fee on credit purchases rather than a flat per-token rate. If breadth of model choice is your priority rather than credential handling, it is the stronger pick.

LiteLLM is the open-source option. It proxies more than a hundred model APIs behind an OpenAI-compatible interface, is free to self-host, and gives you complete control over routing and logging. You pay in operational work: you run it, you scale it, you get paged when it breaks. Teams with platform engineers often prefer this.

Cloudflare AI Gateway is the closest architectural match, another networking company applying the same idea, with caching, rate limiting and analytics on top and a free tier to start. If your stack already sits on Cloudflare, it is the path of least resistance for the same reason ngrok’s gateway is if you already use ngrok.

Final verdict: is ngrok AI Gateway worth it?

Yes, if you are routing to more than one provider and you value the credential model. The split between client-facing access keys and server-side provider keys is a real security improvement, not a marketing line, and $0.05 per million tokens is a fair price for the routing, failover and key handling you would otherwise maintain yourself. The drop-in SDK compatibility means trying it costs an afternoon.

Be clear-eyed about the two limits. The built-in provider list is short, so anything beyond OpenAI and Anthropic means bringing your own keys, and the bundled inference pricing hides your model costs at exactly the moment you would most like to see them. Start on built-in inference to get moving, then switch to your own keys before your volume matters. On that plan, this is a solid piece of infrastructure from a company that understands the layer it is operating at. Anyone calling a single provider should not bother.