FastRouter.ai launched on Product Hunt with a one-line pitch: route requests to the right LLM for cost, latency, and quality. That is the whole product description. No benchmark table, no funding announcement, no named customers. Just a router and a promise that picking the right model per request beats picking one model for everything.

The pitch is unglamorous and, for that reason, worth taking seriously. The interesting question is not whether FastRouter works. It is whether the layer it occupies, model routing, becomes a durable business or a feature that every inference provider absorbs into its own platform within eighteen months.

The arbitrage is real, and shrinking

Model routing exists because model prices diverge wildly and keep moving. A frontier model can cost an order of magnitude more per million tokens than a mid-tier model that handles the same classification, extraction, or summarization task acceptably. If a router can send the easy 70 percent of traffic to the cheap model and reserve the expensive model for the hard 30 percent, the savings are immediate and measurable.

That is a genuine arbitrage. It is also a shrinking one. Every major lab has spent the past two years cutting prices and shipping smaller, cheaper variants of its flagship models. When the price gap between tiers narrows, the absolute savings from routing narrow with it. A router that saves a team 40 percent today may save 15 percent in a year, and the router’s own take rate has to survive that compression.

There is a second problem. Routing quality depends on knowing which model is actually better for a given request, and that knowledge is expensive to acquire. You need evaluation infrastructure, a way to score outputs, and a feedback loop that does not require a human to grade every response. FastRouter’s Product Hunt page does not say how it decides. That omission matters more than the launch copy.

What the router has to get right

Three mechanisms decide whether a routing layer is useful or decorative.

First, classification. The router has to infer, before spending money, whether a request is easy or hard. Get this wrong in the cheap direction and quality collapses. Get it wrong in the expensive direction and you have built an expensive proxy.

Second, fallback. Production traffic hits rate limits, outages, and degraded regions. A router that can fail over from one provider to another is genuinely valuable, because it turns a multi-vendor integration project into a config change. This is arguably a stronger pitch than cost savings, and it is the one FastRouter leads with second.

Third, observability. Teams need to see what was routed where, what it cost, and whether quality held. Without that, no engineering lead signs off on sending production traffic through an opaque middle layer.

None of these are novel. LiteLLM, OpenRouter, Portkey, and a cluster of gateway projects already occupy adjacent ground. Cloudflare and the major clouds ship their own routing and gateway primitives. The category is crowded, and the differentiation between players is thinner than any of their landing pages suggest.

Where the margin actually lives

A pure router is a thin proxy. Thin proxies get commoditized. The companies that survive in this layer do it by owning something the model providers do not: the evaluation data, the enterprise contracts, or the operational trust that comes from being the thing that stays up when three providers go down at once.

There is a real business in that. Enterprise buyers will pay for a single API surface across vendors, predictable spend, and a dashboard that finance can read. That is a procurement problem more than a machine-learning problem, and procurement problems have stickier customers than benchmark wins do.

The risk is that the model providers themselves decide routing is their job. Several already expose automatic model selection within their own families. If a customer is willing to stay inside one vendor, the cross-vendor router loses its reason to exist. FastRouter’s entire value proposition depends on customers wanting to stay multi-vendor, which is a bet on vendor distrust as much as on cost.

That bet is not crazy. Teams that have been burned by a single provider’s outage, price change, or deprecation schedule tend to stay multi-vendor. But it is a bet, and the launch page does not make it explicitly.

A router’s moat is not the routing. It is the evaluation data and the uptime record that make routing trustworthy.

What to watch

The signals that matter here are not on the Product Hunt page. Watch whether FastRouter publishes its routing methodology and evaluation harness. Watch whether it names production customers, which would tell you the product survives contact with real traffic. Watch whether it prices per request, per token routed, or as a flat platform fee. Each choice implies a different theory of where the value sits.

The broader read for AI builders is that the stack is thickening at the inference layer. Model choice used to be a decision made once at project start. It is becoming a runtime decision made per request, and that shift creates room for infrastructure that did not need to exist two years ago. Whether FastRouter is the company that captures it is a separate question from whether the layer is real.

The layer is real. The launch page, so far, is one line and a link.