The load-bearing part is the K-nearest-neighbour fit over embedded requests [5]. Such a fit answers one question, which is which model previously handled work that looked like this. The value of the answer scales with how much of your traffic has a close neighbour in your own trace history. An agent hitting the same six tools all day has dense coverage. An agent taking tasks from users has whatever its tail looks like, and the tail is where the expensive errors live.
Which is why the confidence gates are the part I would point at in review. When the evidence is weak, or a request looks unlike the router's training examples, it stays on the strongest known model [7]. That is the correct default, and it fixes the failure mode: unfamiliar traffic costs full price rather than producing bad output. Read the savings figures accordingly, as a measurement of how repetitive the traffic being routed was.
Upstream of the router sits the piece that is harder to verify. World Model Optimizer ingests traces and constructs text simulations of the tools and environments an agent encounters, then runs candidate models against those simulated tasks [4]. A cheaper model is therefore promoted on evidence gathered against a reconstruction of your API rather than your API. Anywhere the reconstruction is more forgiving than production, the routing table inherits the error. Fallah led development of a mixed-reality simulator at Waabi for testing autonomous vehicles [8], so simulator fidelity is a problem this team has already met under harsher conditions.
Experiential Labs says plainly that the published numbers are configurations it chose and ran itself [14]. Their spread is the interesting bit. RouterBench came in 66.5% cheaper for a 1.7% performance drop [11], which leaves 33.5% of baseline spend, about a third [1]. TauBench came in 44.5% cheaper with performance 6.3% higher [12]. Terminal Bench 2 came in 64% cheaper [13]. Elsewhere the company advertises up to 97% lower cost on some tasks [17], which sits 30.5 points beyond its own best published benchmark [3], and a guaranteed 50% saving at equal or better quality [17] that its own TauBench figure misses by 5.5 points [2]. The guarantee reads as commercial terms rather than a result.
For any of those figures to move to your workload you would need a comparable candidate model set, a comparable task mix, and comparable prompt cache hit rates, since the router prices caching into its estimate of what a request effectively costs [6]. Change the cache economics and the cheap option changes with them.
The adoption cost is legible. You run the gateway locally, connect provider keys, and build the router from OpenTelemetry traces [10], which means instrumenting the agent and retaining traces in a form the optimizer can read. That retained corpus is the actual bet, not the routing table. Experiential Labs says customers will own the weights its training workflow produces [16], and a commenter on the launch thread asked for precisely the artefact that would test it: a model trained with the method, shown against its base [15]. Until that exists, this is a cost control with an option attached, and the option is denominated in traces you have to start keeping before you know what they are worth.
