Benchmark 01 · Financial Analysis & Research · v1.0
Diligence Stack Agent
Which model delivers the strongest average performance at the best estimated cost per task?
We built a common Diligence Stack agent harness and made only the minimal model-specific changes needed to optimize each model for it. We then ran Financial Modeling and Financial Health & Research tasks through that system to compare average performance, estimated cost per task, and the tradeoff between them.
Average model performance
Model leaderboard
Financial Modeling + Financial Health & Research
Total score by model
Results compare average quality with estimated cost per task.
Subcategory performance
Normalized rubric scores averaged across evaluated work
| Model | Completion | Accuracy | Sources | Analysis | Artifacts | Audit & process | Efficiency |
|---|---|---|---|---|---|---|---|
| GPT 5.6 Sol xhighGPT 5.6 | 94.8 | 87.0 | 90.2 | 97.0 | 94.5 | 67.2 | 41.3 |
| GPT 5.6 Luna xhighGPT 5.6 | 92.0 | 82.0 | 90.8 | 89.4 | 83.0 | 65.0 | 91.2 |
| Claude 5 Fable HighClaude 5 | 90.5 | 77.0 | 89.1 | 89.7 | 81.4 | 86.6 | 38.3 |
| Claude 5 SonnetClaude 5 | 88.8 | 73.0 | 91.5 | 90.7 | 80.4 | 84.6 | 51.5 |
| GPT 5.5 highGPT 5.5 | 92.8 | 84.0 | 88.3 | 85.4 | 75.0 | 55.1 | 58.0 |
| GPT 5.6 Terra xhighGPT 5.6 | 86.5 | 83.0 | 82.0 | 85.0 | 74.4 | 63.4 | 80.0 |
| Grok 4.5Grok | 89.5 | 70.0 | 80.0 | 88.0 | 71.0 | 73.0 | 77.0 |
| Kimi K3Kimi | 90.0 | 68.0 | 80.0 | 88.0 | 83.0 | 80.0 | 30.0 |
| Claude 4.8 Opus highClaude 4.8 | 88.0 | 66.0 | 80.8 | 82.4 | 73.4 | 83.6 | 66.0 |
| GPT 5.6 Terra highGPT 5.6 | 80.0 | 74.0 | 81.7 | 82.0 | 72.7 | 63.3 | 82.9 |
| Claude 5 Fable High (Claude Cowork)Claude 5 | 80.0 | 64.0 | 80.0 | 80.0 | 70.0 | 63.3 | 8.6 |
| Muse Spark 1.1Muse | 85.0 | 58.0 | 82.5 | 83.3 | 48.0 | 50.0 | 30.0 |
| GPT 5.6 Terra mediumGPT 5.6 | 72.0 | 64.0 | 76.0 | 74.0 | 62.7 | 58.9 | 80.0 |
| DeepSeek V4 ProDeepSeek | 80.0 | 53.0 | 57.8 | 81.0 | 61.9 | 51.0 | 80.0 |
| Gemini 3.5 Flash HighGemini | 76.8 | 56.0 | 58.8 | 75.0 | 67.5 | 50.5 | 62.0 |
| Kimi K2.6Kimi | 75.0 | 38.0 | 55.0 | 56.7 | 68.0 | 60.0 | 80.0 |
| Grok 4.3Grok | 40.0 | 42.0 | 65.0 | 56.7 | 0.0 | 30.0 | 0.0 |
Why we built this benchmark
Diligence work is not a single prompt. This benchmark asks which model performs best, and at what cost, across multiple realistic tasks inside the same agent system.
We do not expect an agent's output to represent the end of a workflow. For most users, it will be a first draft that still benefits from expert review, additional research, and refinement.
The goal is to measure the quality of that starting point. The more complete, accurate, well-sourced, and usable the first draft is, the more work it can eliminate downstream.
This benchmark therefore evaluates more than whether a model can produce a plausible answer. It examines whether the complete agent system can find the right evidence, use the available knowledge and tools, reason across sources, and turn that research into coherent, review-ready work across the evaluated workflow.
- Objective
- Strongest average first draft
- Total score
- Average of available evaluations
- Estimated cost
- Average cost per task
Examples of tasks we evaluate.
These examples illustrate the practical diligence work in the benchmark; they are not an exhaustive list of the workflows the system can evaluate.
Financial Modeling
Research-backed financial model, decision-ready PDF, and auditable workbook.
- Rubric
- Specific
- Score scale
- 0–100
Financial Health & Research
Financial-health assessment, source-grounded analysis, PDF, and workbook.
- Rubric
- Specific
- Score scale
- 0–100
And more!
What each evaluation measures
Category names and weights vary by workflow, while the rubrics evaluate the same core qualities of useful diligence work.
Completion
Task-level rubricFollows the requested scope, format, and decision criteria.
Accuracy
Task-level rubricGets the material facts right and avoids unsupported claims.
Source grounding
Task-level rubricUses relevant evidence with clear attribution and traceability.
Analysis
Task-level rubricTurns evidence into useful, well-supported judgment.
Artifact quality
Task-level rubricProduces readable, usable, and reviewable files when required.
Auditability & process
Task-level rubricPreserves provenance, checks, tool reliability, and a defensible trail.
Efficiency
Task-level rubricBalances output quality with the cost of producing the result.
Every task rubric resolves to the same 0–100 scale before averaging.
A purpose-built knowledge environment
The benchmark does not rely exclusively on public web search or a generic document store. Agents work with two private, purpose-built knowledge bases assembled for Diligence Stack research workflows.
- Professional grade investment research reports
- 1,000+
- Creative Strategies and Diligence Stack reports
- ~100
The underlying documents are largely unstructured. Agentic preprocessing is designed to identify and preserve document hierarchy, companies, industries, dates, tables, financial metrics, forecasts, themes, and source metadata.
The processed material is indexed with hybrid retrieval that combines semantic, lexical, and metadata-aware search. This is designed to improve both conceptual source discovery and exact recovery of names, figures, dates, product references, and industry terminology.
In our internal testing, this approach materially outperformed conventional retrieval options in source relevance, information recovery, and the agent's ability to locate the right research for a task.
Why the knowledge base matters
Public sources complement the provided research for current developments, primary-source verification, and genuine gaps. The benchmark evaluates whether an agent:
- Searches the appropriate knowledge base
- Selects relevant reports instead of retrieving volume for its own sake
- Uses professional grade investment research and proprietary team analysis appropriately
- Preserves source attribution and traceability
- Distinguishes source evidence from its own inference
- Uses public search deliberately to verify or fill genuine gaps
- Synthesizes internal and external sources into a useful result
How Diligence Stack Agent Bench works
Benchmark 01 · v1.0 · Financial Analysis & Research
For each model, total score is the arithmetic mean of available evaluation scores. Estimated cost per task is the arithmetic mean of reported costs.
- 01
Representative work
The benchmark includes practical diligence work such as Financial Modeling and Financial Health & Research.
- 02
Common agent harness
The core Diligence Stack harness stays consistent, with only minimal model-specific optimization.
- 03
Task-level grading
Each completed evaluation receives a 0–100 score under its task-specific rubric, with raw cost recorded separately.
- 04
Model-level averages
Total score is the mean of all available scores; estimated cost per task is the mean of all reported evaluation costs.
What is actually being tested
The benchmark evaluates an agent configuration, not a base model in isolation. Each result reflects the complete system that turns a task into sourced research, structured analysis, and usable supporting work.
- 01Underlying model
- 02Reasoning configuration
- 03Agent harness
- 04System instructions
- 05Tools and MCP integrations
- 06Diligence Stack skills
- 07Knowledge-base access and retrieval behavior
- 08Task-specific analysis and artifact workflow
This is as much a test of how a model performs inside the Diligence Stack harness as it is a test of the model itself. Tools, retrieval, skills, source access, instructions, and artifact workflow all shape the result.
Every score should therefore be understood as a specific model-and-harness configuration. The same model may perform very differently when its tools, retrieval system, skills, instructions, or artifact workflow change.
How to read this result
The objective is not a perfect final deliverable. It is the strongest and most defensible first draft an agent can produce before a human expert takes over.
Arithmetic means
Averages summarize the available evaluations and do not imply a statistical confidence interval.
Evolving coverage
The published benchmark will broaden as additional representative diligence work is evaluated.
Task-specific rubrics
Tasks retain their own weighted rubrics; only their final 0–100 scores are combined in the model average.
System-level result
Scores reflect the model, harness, instructions, tools, retrieval, and artifact workflow operating together.
Human-guided calibration
Judge results are calibrated and validated against our evaluation guidelines.
Compare quality and cost across the complete agent workflow.
Review the results, averaging method, knowledge environment, and system boundaries behind Diligence Stack Agent v1.0.