Baseten Router nears Opus 5.5 at 72% lower cost
Baseten has published first benchmark results for its coding-task router, which comes within 0.8 points of Opus 5.5's pass rate while cutting costs 72% by sending easier tasks to a cheaper open-weight model.
Baseten published initial results for Baseten Router, the model router that is part of its Baseten Grid platform. We covered the router's launch earlier, and Grid itself launched as a way to cut AI coding costs. Both are now in public preview.
The router reads the input prompt and decides whether a closed frontier model or a cheaper open-weight model should handle the task. In Baseten's experiments, it chose between Opus 5.5 and GLM 5.3 Flash, which Baseten says is about 80% cheaper than Opus 5.5 on relevant workloads. The router gives each model a score: its predicted success rate minus a penalty based on its predicted cost. A configurable weight sets the trade-off, so users can choose different points on the cost-performance Pareto frontier.
The test set was DeepSWE 1.1, SWE Atlas QnA and Terminal-Bench 2.1. On it, the router came within 0.8 percentage points of Opus 5.5's pass rate at 72% lower cost than running Opus 5.5 on every task. Baseten says it was the only router it tested that beat random routing by a statistically significant margin. The company did not name the other routers in the material reviewed here.
Training data came from Opus 5.5 and GLM 5.3 Flash runs on public coding datasets, using high reasoning effort. Baseten says it fixed grader errors and broken task environments, and removed training tasks whose repositories also appeared in the test set. It also tried to account for noise from non-deterministic agent runs.
All of these are benchmark results. Baseten says it is working with design partners on production coding workloads and will share those results later.