Ramp Labs

Ramp Accounting Bench

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.

Best reported

61.1

Gemini 4 ArgonGoogle

mean reward % (criteria, mean of 3 attempts)

Official board, 14 Oct 2026

About this benchmark

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.

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: Gemini 4 Argon, 61.1 at $3.03 per question (cheapest within 5% of the top score)

Ramp Accounting Bench: score vs cost

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

Updated 31 Oct 2026

Y: mean reward % (criteria, mean of 3 attempts)X: Cost per test (USD)Showing 5 models
Ramp Accounting Bench: score vs cost5 models. Frontier: Claude Opus 5.5, Gemini 4 Argon.505560$0$1$2$3EfficientInefficient↑ Best performance← Lowest costCost per test (USD)Gemini 4 Argon: 61.1 mean reward % (criteria, mean of 3 attempts), $3.03 per test (Ramp Labs Accounting Bench, 14 Oct 2026)Gemini 4 ArgonClaude Opus 5.5: 50.5 mean reward % (criteria, mean of 3 attempts), $1.34 per test (Ramp Labs Accounting Bench, 21 Sept 2026)Claude Opus 5.5Claude Fable 5.1: 49.7 mean reward % (criteria, mean of 3 attempts), $3.48 per test (Ramp Labs Accounting Bench, 2 Sept 2026)Claude Fable 5.1GPT-6 Astra: 48.7 mean reward % (criteria, mean of 3 attempts), $2.31 per test (Ramp Labs Accounting Bench, 5 Sept 2026) GPT-6 AstraMuse Wonder: 48.7 mean reward % (criteria, mean of 3 attempts), $2.65 per test (Ramp Labs Accounting Bench, 31 Oct 2026)Muse WonderRamp Accounting Bench: score vs cost5 models. Frontier: Claude Opus 5.5, Gemini 4 Argon.505560$1$2$5EfficientInefficient↑ Best performance← Lowest costCost per test (USD)Gemini 4 Argon: 61.1 mean reward % (criteria, mean of 3 attempts), $3.03 per test (Ramp Labs Accounting Bench, 14 Oct 2026)Gemini 4 ArgonClaude Opus 5.5: 50.5 mean reward % (criteria, mean of 3 attempts), $1.34 per test (Ramp Labs Accounting Bench, 21 Sept 2026)Claude Opus 5.5Claude Fable 5.1: 49.7 mean reward % (criteria, mean of 3 attempts), $3.48 per test (Ramp Labs Accounting Bench, 2 Sept 2026)Claude Fable 5.1GPT-6 Astra: 48.7 mean reward % (criteria, mean of 3 attempts), $2.31 per test (Ramp Labs Accounting Bench, 5 Sept 2026) GPT-6 AstraMuse Wonder: 48.7 mean reward % (criteria, mean of 3 attempts), $2.65 per test (Ramp Labs Accounting Bench, 31 Oct 2026)Muse Wonder

Ramp Accounting Bench

mean reward % (criteria, mean of 3 attempts), higher is better. 5 models.

Updated 31 Oct 2026

  1. Gemini 4 Argon61.1
  2. Claude Opus 5.550.5
  3. Claude Fable 5.149.7
  4. GPT-6 Astra48.7
  5. Muse Wonder48.7

Ramp Accounting Bench: cost

cost per test item (USD), lower is better. 5 models.

Updated 31 Oct 2026

  1. Claude Opus 5.51.3
  2. GPT-6 Astra2.3
  3. Muse Wonder2.6
  4. Gemini 4 Argon3.0
  5. Claude Fable 5.13.5

Ramp Accounting Bench

mean reward % (criteria, mean of 3 attempts), higher is better. 5 models.

Updated 31 Oct 2026

  1. 1Gemini 4 Argon61.1
  2. 2Claude Opus 5.550.5
  3. 3Claude Fable 5.149.7
  4. 4GPT-6 Astra48.7
  5. 4Muse Wonder48.7
Ramp Accounting Bench leaderboard (mean reward % (criteria, mean of 3 attempts), higher is better)
1Gemini 4 ArgonGoogle61.1137 tasks, ReAct harness, GPT 5.6 Luna judgeRamp Labs Accounting Bench
Official board, reported 14 Oct 2026 · checked 2 Nov 2026
2Claude Opus 5.5Anthropic50.5137 tasks, ReAct harness, GPT 5.6 Luna judgeRamp Labs Accounting Bench
Official board, reported 21 Sept 2026 · checked 2 Nov 2026
3Claude Fable 5.1Anthropic49.7137 tasks, ReAct harness, GPT 5.6 Luna judgeRamp Labs Accounting Bench
Official board, reported 2 Sept 2026 · checked 2 Nov 2026
4GPT-6 AstraOpenAI48.7137 tasks, ReAct harness, GPT 5.6 Luna judgeRamp Labs Accounting Bench
Official board, reported 5 Sept 2026 · checked 2 Nov 2026
4Muse WonderMeta48.7137 tasks, ReAct harness, GPT 5.6 Luna judgeRamp Labs Accounting Bench
Official board, reported 31 Oct 2026 · checked 2 Nov 2026

Last checked 2 Nov 2026. Something out of date or wrong? Send a correction.