AgenticFinBench - Data and Calculation: leaderboard
Financial agent research on230 real-world questions with eight categories and four complexity levels. Published quality score Q (0–100). AgenticFinBench230questions, native expert-defined quality score Q in0–100; score-table abbreviated labels bound to exact Table14 model/harness/effort/finance-tool configuration and Table15 execution controls. Most runs use2100s attempts; hosted client exceptions remain as published. Different harnesses are distinct source configurations, never pooled as one model run. Cohort includes Seed2.1Pro0915 and opaque DuMate as a named system, not an assumed underlying model.
Metric: Published quality score Q (0–100). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol (Benchmark Sweep Run: source=AgenticFinBench; harness=Codex CLI; effort=high; finance-tools=✓; tool-scope=Web search / page access; command execution) | 95.71 | |
| 2 | GPT-5.6 Sol (Benchmark Sweep Run: source=AgenticFinBench; harness=Pi 0.87.1; effort=high; finance-tools=✗; tool-scope=Search, page retrieval, shell, and file tools) | 94.28 | |
| 3 | Seed-2.1-pro 0915 (Benchmark Sweep Run: source=AgenticFinBench; harness=Trae Work; effort=—; finance-tools=✓; tool-scope=Wind skill and data-access captures) | 93.82 | |
| 4 | Gemini 4 Argon (Benchmark Sweep Run: source=AgenticFinBench; harness=Gemini CLI 0.64.0; effort=high; finance-tools=✗; tool-scope=google_web_search, web_fetch) | 93.78 | |
| 5 | DeepSeek V4.1 Flash (Benchmark Sweep Run: source=AgenticFinBench; harness=Pi; effort=high; finance-tools=✗; tool-scope=read, bash, edit, write, web_search, web_fetch) | 93.72 | |
| 6 | DeepSeek V4.1 Flash (Benchmark Sweep Run: source=AgenticFinBench; harness=DSH; effort=high; finance-tools=✗; tool-scope=web_search, web_fetch; shell, file, image, planning, and workflow tools) | 93.65 | |
| 7 | GPT-6 Astra (Benchmark Sweep Run: source=AgenticFinBench; harness=Codex CLI; effort=medium; finance-tools=✓; tool-scope=Web search / page access; command execution) | 92.74 | |
| 8 | DeepSeek V4.1 Flash (Benchmark Sweep Run: source=AgenticFinBench; harness=Codex CLI 0.162.0-alpha.2; effort=high; finance-tools=✗; tool-scope=Search, page retrieval, command and file tools) | 92 | |
| 9 | Claude Opus 5.5 (Benchmark Sweep Run: source=AgenticFinBench; harness=Codex CLI 0.162.0-alpha.2; effort=medium; finance-tools=✗; tool-scope=Responses-to-Anthropic protocol adapter; recorded native tool calls) | 92 | |
| 10 | Kimi K3 (Benchmark Sweep Run: source=AgenticFinBench; harness=Kimi hosted + finance suite; effort=high; finance-tools=✓; tool-scope=Finance datasource skills; web search, Python, and artifact tools) | 91.94 |
Interactive version: theaggregate.ai/benchmark?slug=agenticfinbench-data-and-calculation · How It Works · Data refreshed daily, snapshot 2026-11-06.