Workflow evals

Four: security incidents, agent trace observability, invoice processing, and customer service. Models are scored as workflows of Noul, Choice, and Score questions against consensus labels from GPT-6 Astra and Claude Fable 5.1 at high thinking.

As vendor-run evals: Jev averages 67.8% accuracy at $0.0004 and 0.4s per case across the four workflows, and every model is more accurate, cheaper, and faster in a workflow than with the same policy as a prompt, on average. The harness is assumed correct, not independently audited.

It is the page's headline claim: real-world tasks can run as structured workflows or standalone prompts, and the charts show every model is more accurate in a workflow than with the same policy as a prompt, on average.

One chart plots accuracy against cost per case and the other plots accuracy against seconds per case. Both use logarithmic scales, with up and to the left marked better.

That nothing is both cheaper and more accurate, with Jev on the frontier at 67.8% for $0.0004 per case, against for example opus 5 at 73.1% for $0.1761.

Each model's workflow row is more accurate, cheaper, and faster than its prompt row on average. Haiku 4.5 moves from 18.1% at $0.0363 and 21.2s as a prompt to 53.6% at $0.0195 and 12.5s as a workflow.

Jev is shown as a workflow at 67.8% accuracy, $0.0004 per case, and 0.4 seconds, which the page marks as the frontier where nothing is both faster and more accurate.