Step 5 Preview Tops M&A Memo Value Test

The compelling story here is not raw benchmark supremacy but useful financial work at lower cost. Still, a six-model contest is an encouraging signal—not a definitive replacement for analyst-grade evaluation.

• –Cost-per-useful-output is a better metric for finance agents than isolated intelligence scores.
• –Processing 80 companies demonstrates how agents could compress research workflows that traditionally require large analyst teams.
• –Two AI judges introduce evaluator bias; human review, citation accuracy, and factual consistency still need testing.
• –Step 5 Preview’s 1M-token context and tool-use focus fit large filings, market feeds, and multi-step diligence workflows.
• –“Workstation” deployment needs qualification: official serving guidance targets multi-GPU systems, including four B300 GPUs for FP8. GitHub.