Julius Research's DeepFinance benchmark: the best model scores 54.8% on 27-hour private-equity and banking tasks
The 21 tasks, built from anonymised deal data and graded on 84 criteria each, put Claude Opus 5.5 first at 54.8%, just ahead of Claude Sonnet 5.5 (54.6%) and GPT-6 Astra (54.2%); the scores are Julius's own run.
Why it matters
DeepFinance asks models to do end-to-end work in private equity and investment banking, which Julius estimates would take a professional 20 to 30 hours per task, 27.1 on average. Graders score workflow and judgment, such as whether assumptions are sourced and auditable, as well as the final output. Julius says it ran each vendor's own command-line tool rather than its harness. Cost varies far more than score: GPT-6.1 Sol reached 52.9% at $0.81 a task, against $8.63 for the leader. All figures come from Julius Research, the lab behind the Julius data-analysis product.
Key facts
- 21 anonymised private-equity and investment-banking tasks
- Average 27.1 human hours per task; 84 rubric criteria each
- Claude Opus 5.5 (max): 54.8% ± 9.1 at $8.63 per task
- Claude Sonnet 5.5 54.6%; GPT-6 Astra 54.2%
- GPT-6.1 Sol: 52.9% at $0.81 per task
- 16 model configurations, scored in Julius's own run