DeepFinance

DeepFinance is an ultra-long-horizon finance-agent benchmark from San Francisco lab Julius Research (makers of Julius, the AI data scientist with 2M+ users; used by Harvard Business School, Nvidia and others), shipped 2026-11-05. It measures models on real Private Equity and Investment Banking workflows. Site: https://www.juliusresearch.ai/deep-finance .

Official scale:

• 21 real-deal tasks;
• ~27.1 human hours per task on average;
• ~84 rubric criteria per task;
• ~19 files per data room;
• 10 attempts per task with mean rubric % and 95% bootstrap CIs.

Leaderboard (updated Nov 5, 2026): Claude Opus 5.5 (max) 54.8%, Claude Sonnet 5.5 54.6%, GPT-6 Astra 54.2%, GPT-6.1 Sol 52.9% — most frontier configs still far from human-completable. The site also plots score vs. cost/tokens with a Pareto frontier.

Why it matters

Most finance/agent benches are short-form or multiple-choice. DeepFinance stretches into multi-dozen-hour real deal workflows with fine-grained rubrics instead of a single pass/fail — closer to what PE/IB associates actually ship.

Positioning

Not general finance Q&A or a short trading sim: the focus is data rooms, long-horizon tool use and auditable criteria. Cost and token columns make Pareto / value selection practical.

Limits

• Scores come from the hoster's harness; independent reproduction is expensive;
• Leaders sit near ~55% — ceilings and rubric calibration will keep moving;
• PE/IB coverage does not generalize to all finance work.

Who it's for

• Teams building finance agents or long-horizon agent evals;
• People picking models on quality-vs-cost using the official board;
• Anyone following Julius Research's eval line.

Getting started

1. Read https://www.juliusresearch.ai/deep-finance for task design and the leaderboard;
2. Use the score/cost/tokens table to shortlist configs;
3. Note the update stamp (Updated Nov 5, 2026);
4. If you rebuild locally, match rubrics and attempt counts before comparing numbers.