# Cortex Router on tau2-bench: Open Models

## Headline

On the 99-task common cohort, the Cortex task-level router reaches **81% task success** at **$0.159 per task**. Against Kimi K3, it changes success by **-6 percentage points** and reduces mean cost by **61%**.

## Results

The table reports all valid rows available for each arm; denominators are shown explicitly.

| Configuration | Passed | Task success | Cost / task |
|---|---:|---:|---:|
| Qwen3.7 Plus | 20/100 | 20% | $0.103 |
| GLM-5.2 | 83/100 | 83% | $0.307 |
| Kimi K3 | 86/99 | 87% | $0.413 |
| Cortex task-level router | 81/100 | 81% | $0.159 |

## Charts (sticky source)

![Cost vs quality](./cost_quality_chart.png)

![Operating distributions](./operating_distributions.png)

![Paired task outcomes](./paired_task_outcomes.png)

![Router mix](./router_mix.png)

## Method

Retrospective task-level routing estimate over a 100-task tau2-bench base-split pilot. One model is selected from the initial task prompt and retained for every turn. Infrastructure failures are excluded from the affected fixed-arm denominator. Headline comparisons use only the common valid task IDs.

Per-task records: [CSV](/cortex/benchmarks/tau2-open-model-router/per_task_results.csv) and [JSON](/cortex/benchmarks/tau2-open-model-router/per_task_results.json). Claim map: [JSON](/cortex/benchmarks/tau2-open-model-router/claims.json). [Recomputation notes](/cortex/benchmarks/tau2-open-model-router/recompute.md).