Liner, the AI agent solutions company, has launched Liner Model API — a model-routing API designed to cut enterprise LLM costs by automatically sending each request to the most cost-efficient model capable of handling it, rather than defaulting to an expensive frontier model for everything.
The pitch targets a real inefficiency: companies routinely pay flagship-model prices for AI tasks — simple questions, routine classification, light reasoning — that cheaper models handle just as well. Liner's system evaluates the expected quality and cost of candidate models per request and routes it to the cheapest one that clears the bar, from everyday questions to demanding coding, reasoning, and deep-research workloads. Crucially, it picks one model per request rather than calling several at once, avoiding the extra token spend of multi-model approaches.
For credibility, Liner points at its own bills: after deploying its Liner Orchestrator internally, the company says its token expenses in August dropped by more than 50% compared with the first half of 2026. The developer pitch is simplicity — one API across workloads that would otherwise need ongoing model selection and cost optimization. It is the same cost-pressure logic behind this week's wave of cheaper models (Claude Opus 5.5, GPT-6 Sol and Luna): the frontier race is now about price-performance per token, not just raw capability.
The insight behind model routing is almost embarrassingly obvious once stated: not every AI task needs a frontier model. Yet most companies default to exactly that — the most expensive model for everything — because choosing per task is operational work nobody wants to do. It is the cloud-computing story retold: enterprises overprovisioned servers for years until tooling made right-sizing automatic. Liner is betting the same laziness tax exists in AI spend, and its own 50% internal savings suggest the tax is large.
What is clever about the approach is the single-model-per-request design. The naive way to route — try several models and pick the best answer — burns the very tokens you are trying to save. Predicting quality before inference is harder engineering, but it is the only version of routing whose economics actually work. If Liner's quality predictions hold up across diverse workloads, it has a real moat; if they do not, customers will notice wrong answers long before they notice the savings.
Zoom out and this launch is another marker of the industry's phase change: the frontier race is now fought on price-performance per token, not just capability. When Claude, GPT, and open models all clear the "good enough" bar for most tasks, the winner is whoever delivers that quality cheapest. Model routers, cheaper models, and aggressive API pricing are all the same trend wearing different clothes — the commoditization of intelligence, arriving right on schedule.
The endgame is obvious: routing becomes table stakes, built into every AI platform, and the standalone router either gets acquired or becomes infrastructure. Until then, there is a window where smart routing is pure margin — and Liner just jumped through it.