Atlas

Benchmarks

← All benchmarks

DeepFinance

Finance & Accounting · 2026-11-05

DeepFinance (Julius Research) tests agents on 21 private equity and investment banking tasks in valuation, due diligence, underwriting and financial modeling that would take a professional 20 to 30 hours. Each task pairs a request with a data room of about 19 files, drawn from anonymized real deals and including conflicting sources, and is graded against about 84 expert-written rubric criteria covering the output, the workflow and the judgment behind it. The score is the mean percentage of criteria satisfied over 10 successful attempts per task. Each model runs in the harness its lab ships, with web search on.

Benchmark source: Juliusresearch

Top models (higher is better)

ModelScoreSource
Claude Opus 5.554.8DeepFinance leaderboard: Claude Opus 5.5 max · Julius Research
Claude Sonnet 5.554.6DeepFinance leaderboard: Claude Sonnet 5.5 max · Julius Research
GPT-6 Astra54.2DeepFinance leaderboard: GPT-6 Astra xhigh · Julius Research
GPT-6.1 Sol52.9DeepFinance leaderboard: GPT-6.1 Sol xhigh · Julius Research
Claude Fable 5.145.6DeepFinance leaderboard: Claude Fable 5.1 max · Julius Research
GPT-6.1 Terra43.5DeepFinance leaderboard: GPT-6.1 Terra xhigh · Julius Research
Gemini 4 Argon41.3DeepFinance leaderboard: Gemini 4 Argon high · Julius Research
Muse Spark 1.337.0DeepFinance leaderboard: Muse Spark 1.3 max · Julius Research
Kimi K3.1 Pro34.6DeepFinance leaderboard: Kimi K3.1 Pro max · Julius Research
Grok 4.734.1DeepFinance leaderboard: Grok 4.7 xhigh · Julius Research
Inkling 2 Small29.8DeepFinance leaderboard: Inkling 2 Small max · Julius Research
GLM-5.327.7DeepFinance leaderboard: GLM 5.3 max · Julius Research
GPT-6 Luna24.6DeepFinance leaderboard: GPT-6 Luna xhigh · Julius Research
DeepSeek-V4-Pro20.5DeepFinance leaderboard: DeepSeek V4 Pro max · Julius Research
Inkling13.1DeepFinance leaderboard: Inkling max · Julius Research
Inkling-Small10.9DeepFinance leaderboard: Inkling Small max · Julius Research
Loading Atlas data…