# Ramp Accounting Bench leaderboard

Long-horizon accounting tasks: agents find evidence across QuickBooks ledgers, spreadsheets and PDFs, apply accounting treatments, and produce exact answers, edited files or new artefacts.

## Ramp Accounting Bench: mean reward % (criteria, mean of 3 attempts) (higher is better, 5 models)

| Rank | Model | Provider | Score | Metric | Setting | Reported by | Source | URL | Date |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| 1 | Gemini 4 Argon | Google | 61.1 | mean reward % (criteria, mean of 3 attempts) | 137 tasks, ReAct harness, GPT 5.6 Luna judge | Official board | Ramp Labs Accounting Bench | https://labs.ramp.com/accountingbench | 14 Oct 2026 (checked 2 Nov 2026) |
| 2 | Claude Opus 5.5 | Anthropic | 50.5 | mean reward % (criteria, mean of 3 attempts) | 137 tasks, ReAct harness, GPT 5.6 Luna judge | Official board | Ramp Labs Accounting Bench | https://labs.ramp.com/accountingbench | 21 Sept 2026 (checked 2 Nov 2026) |
| 3 | Claude Fable 5.1 | Anthropic | 49.7 | mean reward % (criteria, mean of 3 attempts) | 137 tasks, ReAct harness, GPT 5.6 Luna judge | Official board | Ramp Labs Accounting Bench | https://labs.ramp.com/accountingbench | 2 Sept 2026 (checked 2 Nov 2026) |
| 4 | GPT-6 Astra | OpenAI | 48.7 | mean reward % (criteria, mean of 3 attempts) | 137 tasks, ReAct harness, GPT 5.6 Luna judge | Official board | Ramp Labs Accounting Bench | https://labs.ramp.com/accountingbench | 5 Sept 2026 (checked 2 Nov 2026) |
| 4 | Muse Wonder | Meta | 48.7 | mean reward % (criteria, mean of 3 attempts) | 137 tasks, ReAct harness, GPT 5.6 Luna judge | Official board | Ramp Labs Accounting Bench | https://labs.ramp.com/accountingbench | 31 Oct 2026 (checked 2 Nov 2026) |

Scores are never combined across benchmarks. Each row names its metric, source and date.

## About this benchmark

- Maker: Ramp Labs
- Year: Not stated
- Use case: Financial spreadsheets and accounting
- Size: 137 tasks across 22 simulated companies, with 92 domain tools
- Public split: One sample task is shown on the page; the task set is not published
- Grading: Human-written rubric per task, scored criterion by criterion by a frozen LLM judge (GPT 5.6 Luna). Headline is mean reward with partial credit over three attempts; solve rate and pass^3 are also reported.
- Licence: Not stated
- Submission: Not stated
- Products allowed: Not stated
- What it does not cover: No investment research, trading or market data.
- Last checked: 2 Nov 2026
- Links: site https://labs.ramp.com/accountingbench, leaderboard https://labs.ramp.com/accountingbench
- Directory page: https://benchmarks.finance/benchmarks/ramp-accounting-bench/
- Leaderboard: https://benchmarks.finance/leaderboards/ramp-accounting-bench/

Source: https://benchmarks.finance/leaderboards/ramp-accounting-bench/

Created and maintained by Moneda.