Routing is free now. Read the fine print.

In the space of about a month, routing went from a category you had to explain to a category two very large companies decided to give away.

On 19 August Ramp, a corporate card and expense company, launched Router.com, a model gateway with no routing fees through the end of 2026. The next day Google announced model routing inside Cloud API Gateway, in Public Preview, at no extra charge beyond the gateway itself. Both products accept OpenAI-compatible requests. Both promise to pick the model for you.

I run a routing company, so read the rest of this with that in mind. But I would give the same advice about any free infrastructure product: the launch post tells you what it does, and the fine print tells you what it costs. Here is the fine print on both.

Ramp Router: free until it is not

Router.com is a serious product. One endpoint, OpenAI- and Anthropic-compatible APIs, automatic fallback, shadow models, benchmark-driven routing, and what Ramp calls a Switchyard-style escalation pattern where a small model handles most traffic and hands off when it cannot. Ramp claims an average 40% reduction in inference costs for early customers. They also announced Nexus, an agent SDK, at the same time.

The economics, per Ramp: zero routing fee through 2026, users pay each model's list price, and 2027 pricing is to be announced.

Three questions any buyer should be able to answer before moving traffic.

What does the fee become? Ramp's stated model is that the router is a distribution channel. It pays for itself by pulling AI spend onto Ramp cards and into Ramp's expense tooling. That is a coherent strategy, and it might mean routing stays free indefinitely. It might also mean the fee appears in January at whatever level Ramp's finance team decides. Nobody outside Ramp knows, including, I suspect, Ramp. We wrote in free model IDs come with expiry dates about treating promotional inference as a scheduled event. A free router is the same thing one layer up. Put 1 January on the calendar.

Where is the 40% coming from? If Router.com bills at list price and takes no fee, the saving comes from routing to cheaper models. That is exactly what routing is for, and 40% is a credible number for a workload that was sending everything to a flagship. But it is a claim about how badly the customer was routing before, not about the router. Any gateway that moves traffic from a $10 input model to a $1 one with equivalent task success will show the same number. The question to ask is whether your team could see the routing decision and override it per request, or whether it is a black box that reports savings.

Who sees the traffic? Ramp is a fintech with a compliance function and, presumably, a serious security posture. It is also a company whose business model is understanding what its customers spend. Prompts are unusual data to route through an expense platform. Check the retention terms and the data processing agreement before sending anything regulated.

None of that is a reason not to try it. It is a reason to keep the exit cheap.

Google API Gateway: routing within a host

Google's announcement reads like a shift in category. Cloud API Gateway, the product that already fronts REST APIs for many GCP customers, now routes OpenAI-compatible model requests across backends by cost, latency or availability. It is in Public Preview and generally there is no charge for the routing itself beyond the gateway's standard rates.

The constraint is one sentence in the documentation, and it is the sentence that matters. Google's docs state that all backends referenced by a single router must share the same host. In practice, one router covers models served from one Google host. Gemini to Gemini, or Google-hosted open weight model to Google-hosted open weight model. Not Gemini to Anthropic's own API, not Gemini to OpenAI, not Vertex to Bedrock.

That is not a bug in the design. Cloud API Gateway is a GCP product doing what GCP products do: making Google's own surface area easier to consume. Within that scope, the feature is useful. If you are all-in on Vertex, routing between Gemini 3.8 Pro and Gemini 3.8 Flash by cost, with fallback on availability, is worth having and now costs nothing extra.

But it is a different thing from what the phrase "model routing" has come to mean. The reason routing matters in 2026 is that no single provider has the best model for every task, every provider has outages, and prices move between providers on a monthly basis, as we covered in the list price is now a range. A router that can only see one host cannot help with any of that. When Google has a regional incident, as happened in the correlated outage on 3 September, a same-host router fails over to the same failure.

We wrote a longer version of this argument in vendor-agnostic AI gateway options: the cloud gateways are excellent at reducing friction inside their cloud, and that is a different product from reducing your dependence on it.

What both launches tell you about the market

Set aside the fine print for a moment. The signal is that two companies with no history in model infrastructure looked at routing and decided it was worth building and giving away. Ramp wants the spend data and the card volume. Google wants the workloads to stay on Vertex. Both concluded that owning the routing decision is worth more than charging for it.

That tells you three things.

Routing is now table stakes. A year ago you could argue about whether an abstraction layer between your application and your model providers was necessary. When a fintech and a hyperscaler both ship one in the same month, the argument is over. The question is which one.

The routing layer is where leverage sits. Whoever makes the model decision controls where the tokens go, and therefore who gets paid. Stripe understood this when it bought OpenRouter. Ramp understands it now. Google has always understood it. Buyers should understand it too: the router is the most strategic and least visible component in the stack.

Free is a price. Ramp's router is free because it monetises elsewhere. Google's is free because it keeps you on GCP. Neither is wrong. But a free product whose funding comes from somewhere else optimises for that somewhere else, and you should know what it is.

What to check before you move traffic

Reach. List the providers and hosts your workload uses today, including the ones in your fallback chain. Can the router send traffic to all of them? A same-host router is a load balancer with a nice name.

Fee schedule. Get the post-promotion price in writing, or accept that you do not know it and plan for that. If the answer is "to be announced," treat it as a variable cost with unknown magnitude starting 1 January.

Visibility. Can you see which model each request went to, why, and what it cost? Can you override the decision for a request, a key, or a team? A router that reports savings without exposing decisions is asking for trust it has not earned yet. Our own approach is documented in smart routing and routing policies.

Residency. Where are requests processed, where are logs stored, and for how long? For teams under EU obligations, this is the first question, not the last; see EU routing for what the answer should look like.

Exit cost. If any of the above changes, how long does it take to move traffic to another gateway? If your application talks to an OpenAI-compatible endpoint and your routing rules live in configuration rather than code, the answer is an afternoon. If your routing logic lives in the vendor's console, the answer is a migration project.

The takeaway

Two launches in three weeks made routing free and made it mainstream. Both are good for the market. Ramp Router is a capable multi-provider gateway with a funding model that depends on things other than routing fees, and a 2027 price that does not exist yet. Google's API Gateway routing is a solid same-host tool that does not solve the cross-provider problem routing was invented for. Use either if it fits. Just read the fine print first, and keep your routing rules somewhere you can take with you.