# Cortex Router on tau2-bench: GPT-5.6

## Headline

On the 100-task common cohort, the Cortex task-level router reaches **66% task success** at **$0.625 per task**. Against GPT-5.6 Sol, it changes success by **1 percentage point** and reduces mean cost by **52%**.

## Results

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

| Configuration | Passed | Task success | Cost / task |
|---|---:|---:|---:|
| GPT-5.6 Luna | 46/100 | 46% | $0.277 |
| GPT-5.6 Terra | 57/100 | 57% | $0.731 |
| GPT-5.6 Sol | 65/100 | 65% | $1.306 |
| Cortex task-level router | 66/100 | 66% | $0.625 |

## 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-gpt-router/per_task_results.csv) and [JSON](/cortex/benchmarks/tau2-gpt-router/per_task_results.json). Claim map: [JSON](/cortex/benchmarks/tau2-gpt-router/claims.json). [Recomputation notes](/cortex/benchmarks/tau2-gpt-router/recompute.md).