Patronus AI, with Contextual AI and Stanford University, 2023

FinanceBench

Open-book questions about US public-company filings (10-Ks, 10-Qs, 8-Ks, earnings reports). The model must find and compute answers from the filing, usually through retrieval or long context.

Best reported

67

Numa 1Moneda

% correct, no filing given (150-question sample)

Run by Moneda, 2 Nov 2026

About this benchmark

Size
10,231 questions over 361 public filings
Public split
150 annotated questions, with source PDFs and the paper's human-graded model outputs. The full set is available on request from Patronus AI.
Grading
Humans label each answer correct, incorrect or failed to answer. Small unit and rounding slips are allowed.
Licence
CC-BY-NC-4.0 (data, per Hugging Face card)
Submission
via-maker
Products allowed
Not stated
Latest version
arXiv v1 (2023-11-20); HF dataset last modified 2024-11-17
What it does not cover
No live data, no multi-step agent workflows, no non-US filings, and no grading of the reasoning path beyond the final answer.

Results

5 models
Leaderboard

Each score shows who reported it: the benchmark's authors (quoted from their paper), the benchmark's official board, or Moneda (run with the benchmark's own grader and rules). We do not copy other aggregators.

Best value: Numa 1, 67 at $0.041 per question (cheapest within 5% of the top score)

FinanceBench: score vs cost

Score against cost per test. Up and to the left is better.

Updated 2 Nov 2026

Y: % correct, no filing given (150-question sample)X: Cost per test (USD)Showing 3 models
FinanceBench: score vs cost3 models. Frontier: Numa 1.60626466$0$0.05$0.1$0.15EfficientInefficient↑ Best performance← Lowest costCost per test (USD)Numa 1: 67 % correct, no filing given (150-question sample), $0.041 per test (API cost per question (model and search spend), measured in Moneda's run, 2 Nov 2026) Numa 1GPT-5.4: 63.7 % correct, no filing given (150-question sample), $0.044 per test (API cost per question (model and search spend), measured in Moneda's run, 2 Nov 2026) GPT-5.4Claude Opus 5.5: 60.3 % correct, no filing given (150-question sample), $0.156 per test (API cost per question (model and search spend), measured in Moneda's run, 2 Nov 2026)Claude Opus 5.5FinanceBench: score vs cost3 models. Frontier: Numa 1.60626466$0.05$0.1$0.2EfficientInefficient↑ Best performance← Lowest costCost per test (USD)Numa 1: 67 % correct, no filing given (150-question sample), $0.041 per test (API cost per question (model and search spend), measured in Moneda's run, 2 Nov 2026) Numa 1GPT-5.4: 63.7 % correct, no filing given (150-question sample), $0.044 per test (API cost per question (model and search spend), measured in Moneda's run, 2 Nov 2026) GPT-5.4Claude Opus 5.5: 60.3 % correct, no filing given (150-question sample), $0.156 per test (API cost per question (model and search spend), measured in Moneda's run, 2 Nov 2026)Claude Opus 5.5

FinanceBench arXiv v1 (2023-11-20); HF dataset last modified 2024-11-17

% correct, no filing given (150-question sample), higher is better. 5 models.

Updated 2 Nov 2026

  1. Numa 167
  2. GPT-5.463.7
  3. Claude Opus 5.560.3
  4. GPT-4-Turbo19
  5. Llama219

FinanceBench: cost

cost per test item (USD), lower is better. 3 models.

Updated 2 Nov 2026

  1. Numa 10.041
  2. GPT-5.40.044
  3. Claude Opus 5.50.156

FinanceBench arXiv v1 (2023-11-20); HF dataset last modified 2024-11-17

% correct, no filing given (150-question sample), higher is better. 5 models.

Updated 2 Nov 2026

  1. 1Numa 167
  2. 2GPT-5.463.7
  3. 3Claude Opus 5.560.3
  4. 4GPT-4-Turbo19
  5. 4Llama219
FinanceBench leaderboard (% correct, no filing given (150-question sample), higher is better)
1Numa 1Moneda67Finds its own filings, no documents given; graded by a two-judge panel (partly right counts half)Moneda run, 2026-11-02
Run by Moneda, 2 Nov 2026
2GPT-5.4OpenAI63.7Built-in web search, no documents given; same 150 questions and two-judge panel as Numa 1 (partly right counts half)Moneda run, 2026-11-02
Run by Moneda, 2 Nov 2026
3Claude Opus 5.5Anthropic60.3Built-in web search, no documents given; same 150 questions and two-judge panel as Numa 1 (partly right counts half)Moneda run, 2026-11-02
Run by Moneda, 2 Nov 2026
4GPT-4-TurboOpenAI19Shared vector store (all filings)Paper Table 2
Reported by the authors, 20 Nov 2023
4Llama2Meta19Shared vector store (all filings)Paper Table 2
Reported by the authors, 20 Nov 2023

FinanceBench arXiv v1 (2023-11-20); HF dataset last modified 2024-11-17

% correct, the right filing given (150-question sample), higher is better. 3 models.

Updated 20 Nov 2023

  1. GPT-4-Turbo85
  2. Claude276
  3. Llama241

FinanceBench arXiv v1 (2023-11-20); HF dataset last modified 2024-11-17

% correct, the right filing given (150-question sample), higher is better. 3 models.

Updated 20 Nov 2023

  1. 1GPT-4-Turbo85
  2. 2Claude276
  3. 3Llama241
FinanceBench leaderboard (% correct, the right filing given (150-question sample), higher is better)
1GPT-4-TurboOpenAI85Oracle (evidence pages given)Paper Table 2
Reported by the authors, 20 Nov 2023
2Claude2Anthropic76Long context (filing in context)Paper Table 2
Reported by the authors, 20 Nov 2023
3Llama2Meta41Single vector store (one per filing)Paper Table 2
Reported by the authors, 20 Nov 2023

Last checked 1 Nov 2026. Something out of date or wrong? Send a correction.