Liner Details Model API Prices and Claims More Than 50% Internal Token Savings

Liner has published prices and an internal savings result for its model-routing API, which it announced on September 15. The figures sharpen its case for replacing one default AI model with a service that selects a model for each request.

One request, one model

A fixed-model setup sends every request to the same model, including work a cheaper one might handle. Liner Model API instead weighs the expected quality and cost of candidate models, then selects one for each request. It does not call several models at once, a choice Liner says avoids the extra token costs of doing so.

Liner says harder coding, math and reasoning questions go to higher-end models, while simpler ones go to lighter models. In an evaluation of requests from its own services, about 43% went to high-performance models and 57% to more efficient ones without a loss in answer accuracy, the company says. That mix is not a promised split for other customers.

$1Input

Liner lists input at $1 per million tokens.

$6Output

Liner lists output at $6 per million tokens.

$0.10Cached input

Liner lists cached input at $0.10 per million tokens.

Two ways to read the savings claim

Liner reports that after deploying Liner Orchestrator, its August token expenses were more than 50% below its first-half 2026 baseline. Separately, it estimates the API can cost at least 50% less than comparable models in the same performance tier, based on publicly available prices.

Those are different comparisons: one tracks Liner’s expenses across time; the other compares its API with models it considers peers. Neither guarantees a customer’s savings. Liner says actual savings depend on the mix of requests, model use and workload requirements. Customers with many complex questions may see a different result from those with mostly routine ones.

The quality test sits inside the route

A lower bill is useful only if the selected model can do the work. Liner says it refined the router through real-world use and controlled benchmark tests. Its published comparisons include Claude Sonnet 5 and GPT-5.6-Terra. Developers still have to judge whether its model choices meet the quality their own applications require.

The company says it developed its evaluation system while running its own AI services and validated the product using de-identified, representative requests. Developers can create an API key through Liner’s platform, inspect its published benchmarks and use a calculator to estimate savings from their current model spending. The deciding test is whether lower costs come with acceptable answers on their own workloads.