Julius Research

Julius Research is an applied AI research lab based in San Francisco, CA. We’re the creators of Julius (AI Data Scientist with 2M+ users) and used by Harvard Business School and companies like Nvidia.

DeepFinance

An ultra long-horizon benchmark for measuring models on real-world Private Equity and Investment Banking workflows that would take a human professional between 20 and 30 hours to complete.

tasks: 21
average human time: 27.1 hours per task
rubric criteria per task: 84
files per data room: 19

Claude Opus 5.5 (max)54.8% ± 9.1
Claude Sonnet 5.5 (max)54.6% ± 9.0
GPT-6 Astra (xhigh)54.2% ± 8.1
GPT-6.1 Sol (xhigh)52.9% ± 7.8
Claude Fable 5.1 (max)45.6% ± 5.3
GPT-6.1 Terra (xhigh)43.5% ± 6.8
Gemini 4 Argon (high)41.3% ± 7.0
Muse Spark 1.3 (max)37.0% ± 4.6
Kimi K3.1 Pro (max)34.6% ± 4.4
Grok 4.7 (xhigh)34.1% ± 4.4
Qwen 3.8 Max (xhigh)30.3% ± 4.2
GLM 5.3 (max)27.7% ± 3.8
DeepSeek V4.1 Flash (max)27.3% ± 3.9
GPT-6 Luna (xhigh)24.6% ± 4.5
Mistral Large 4 (high)21.5% ± 4.4
DeepSeek V4 Pro (max)20.5% ± 4.1

0%10%20%30%40%50%60%70%

Mean rubric score (%)

Source: DeepFinance (Julius Research)Updated Nov 5, 2026juliusresearch.ai

Results: mean rubric score (%) by model configuration with its 95% confidence interval, best first
Rank	Model	Lab	Mean rubric score with 95% confidence interval
1	Claude Opus 5.5 (max)	Anthropic	54.8% mean rubric score, 95% CI 46.2 to 64.4
2	Claude Sonnet 5.5 (max)	Anthropic	54.6% mean rubric score, 95% CI 46.0 to 63.9
3	GPT-6 Astra (xhigh)	OpenAI	54.2% mean rubric score, 95% CI 46.4 to 62.6
4	GPT-6.1 Sol (xhigh)	OpenAI	52.9% mean rubric score, 95% CI 45.4 to 61.0
5	Claude Fable 5.1 (max)	Anthropic	45.6% mean rubric score, 95% CI 40.7 to 51.3
6	GPT-6.1 Terra (xhigh)	OpenAI	43.5% mean rubric score, 95% CI 37.1 to 50.6
7	Gemini 4 Argon (high)	Google	41.3% mean rubric score, 95% CI 34.7 to 48.8
8	Muse Spark 1.3 (max)	Meta	37.0% mean rubric score, 95% CI 32.6 to 41.9
9	Kimi K3.1 Pro (max)	Moonshot AI	34.6% mean rubric score, 95% CI 30.4 to 39.2
10	Grok 4.7 (xhigh)	SpaceXAI	34.1% mean rubric score, 95% CI 29.9 to 38.7
11	Qwen 3.8 Max (xhigh)	Alibaba	30.3% mean rubric score, 95% CI 26.3 to 34.6
12	GLM 5.3 (max)	Z.ai	27.7% mean rubric score, 95% CI 23.9 to 31.5
13	DeepSeek V4.1 Flash (max)	DeepSeek	27.3% mean rubric score, 95% CI 23.6 to 31.3
14	GPT-6 Luna (xhigh)	OpenAI	24.6% mean rubric score, 95% CI 20.1 to 29.2
15	Mistral Large 4 (high)	Mistral	21.5% mean rubric score, 95% CI 17.2 to 25.9
16	DeepSeek V4 Pro (max)	DeepSeek	20.5% mean rubric score, 95% CI 16.5 to 24.7

Mean rubric score by model configuration on DeepFinance (21 tasks, 10 attempts per task). Rubric score is the percentage of rubric criteria satisfied. Error bars denote 95% bootstrap confidence intervals (100,000 resamples over tasks and attempts).

Score vs. cost

Mean rubric score against mean API cost per task. The dashed line marks the Pareto frontier: configurations not dominated on both score and cost.

Sortable DeepFinance results by model configuration: mean rubric score (higher is better), API cost per task and tokens per task (lower is better)
Rank	Model
1	Claude Opus 5.5 max	54.8, 95% CI 46.2–64.4	$8.63	10.5M
2	Claude Sonnet 5.5 max	54.6, 95% CI 46.0–63.9	$6.94	22.6M
3	GPT-6 Astra xhigh	54.2, 95% CI 46.4–62.6	$4.55	1.72M
4	GPT-6.1 Sol xhigh	52.9, 95% CI 45.4–61.0	$0.81	1.90M
5	Claude Fable 5.1 max	45.6, 95% CI 40.7–51.3	$12.02	3.15M
6	GPT-6.1 Terra xhigh	43.5, 95% CI 37.1–50.6	$0.38	2.15M
7	Gemini 4 Argon high	41.3, 95% CI 34.7–48.8	$2.85	6.69M
8	Muse Spark 1.3 max	37.0, 95% CI 32.6–41.9	$4.87	10.9M
9	Kimi K3.1 Pro max	34.6, 95% CI 30.4–39.2	$2.67	6.08M
10	Grok 4.7 xhigh	34.1, 95% CI 29.9–38.7	$4.71	5.90M
11	Qwen 3.8 Max xhigh	30.3, 95% CI 26.3–34.6	$4.99	12.6M
12	GLM 5.3 max	27.7, 95% CI 23.9–31.5	$1.11	5.67M
13	DeepSeek V4.1 Flash max	27.3, 95% CI 23.6–31.3	$0.13	3.93M
14	GPT-6 Luna xhigh	24.6, 95% CI 20.1–29.2	$0.09	2.80M
15	Mistral Large 4 high	21.5, 95% CI 17.2–25.9	$3.69	5.04M
16	DeepSeek V4 Pro max	20.5, 95% CI 16.5–24.7	$1.50	2.64M

Read the post: DeepFinance: Can AI diligence a $100MM deal?