The 2026 Accounting AI Benchmark

The Accounting AI Benchmark tests leading AI models on 101 tasks across eight areas of accounting, from transaction classification and journal entries to reconciliation, financial reporting, and month-end close. Every model works from the same company books, receives the same instructions, and has access to the same accounting tools. Results are reported across accuracy, consistency, cost, duration, and accounting task area, showing both how models rank and where performance differs.

Current model performance

Compare the latest models across accounting accuracy, cost, consistency, and response time.

Leaderboard

All models in the current round are evaluated using the same benchmark release, accounting environment, tools, instructions, and test conditions.

Key takeaways

Accuracy vs. cost

Headline accuracy does not show how much a result costs to achieve. This view compares accounting performance with the estimated cost of running each model.

Accounting performance by category

Overall accuracy can hide substantial differences between accounting tasks. This view shows how each model performs across the eight areas included in the benchmark.

Gemini 4 Argon leads overall, but Muse Wonder performs better on Bank Reconciliation (91.7% vs. 75.0%) and Kimi K3 on Month-End Close (91.7% vs. 83.3%). Among the top five models, the largest gap appears in Accounts Receivable, where scores range from 50.0% (Kimi K3) to 91.7% (Gemini 4 Argon, Grok 4.7).

Performance by difficulty

The current benchmark contains 37 easy, 36 medium, and 28 hard cases. These labels are assigned as part of the benchmark and should not be read as a universal measure of accounting difficulty.

Gemini 4 Argon, Grok 4.7, and Gemini 3.8 Flash share the strongest performance on hard tasks at 82.1%, while Claude Opus 5.5 and Mistral Large 4 show the largest drop between easy and hard cases (94.6% to 64.3%, a 30.3-point drop).

What the benchmark tests

The Accounting AI Benchmark contains 101 synthetic cases across eight areas of accounting. Each case has a task, a reference answer, and fixed scoring requirements. No customer data is involved.

How answers are scored

Every case has a reference answer written before the model is evaluated. Grading is deterministic: the same response evaluated under the same scoring specification receives the same score. No AI model is used as a judge of another model’s answer.

For classification, structured-response, and numerical cases, a score of 0.8 or higher is a pass. Multiple-choice cases are binary. The benchmark retains both the underlying score and the pass indicator, so a pass should not be read as proof that every accounting detail in the response was perfect.

For numerical cases, expected values are checked using the benchmark’s fixed tolerance rules, including a 0.5% relative tolerance where applicable.

Where we report direct statistical comparisons between models, we compare matched cases using paired bootstrap analysis rather than treating every small percentage-point gap as meaningful.

Benchmark updates

Changes to the benchmark are versioned so readers can see when the case set, scoring rules, evaluation environment, or reporting changed.

October 2026 · Version 1.0

A new task, reference answer, scoring rule, or materially different evaluation condition creates a different benchmark condition. Results from incompatible versions should not be combined into a single ranking.

Methodology

DualBench100 is the evaluation framework behind the Accounting AI Benchmark. It defines the case composition, evaluation procedure, deterministic scoring methods, repeated-trial reporting, cost and latency measurement, matched statistical comparison, and reproducibility requirements.

Every model runs the complete benchmark three times under the same evaluation conditions. Each model receives the same 101 accounting cases, starts from an identical, freshly restored test company, has access to the same accounting tools and permitted actions, and receives the same instructions and evaluation settings. Each result is tied to the exact run, date, model configuration, and Accounting Intelligence release. A task counts as passed when the model passes it in at least two of the three runs.

The benchmark also records execution outcomes, elapsed response time, and available usage data separately from response quality. Cost estimates use provider-reported charges where available or an identified public pricing schedule.

The public benchmark summarizes the parts needed to interpret the results. The full methodology contains the formal scoring definitions, tolerances, eligibility rules, statistical procedure, and interpretation limits.

The DualBench100 methodology

The formal scoring definitions, tolerances, eligibility rules, statistical treatment and interpretation limits behind the Accounting AI Benchmark

Read the paper

How to interpret the benchmark

DualBench100 uses synthetic accounting scenarios across a defined set of tasks. The mix does not represent every company, industry, accounting policy, or workflow, and the easy / medium / hard labels should not be read as universal measures of accounting difficulty.

Scores describe the tested model configuration under the reported conditions. Different instructions, context, tools, settings, or workloads may produce different results.

A benchmark pass means the response met the checks defined for that case. It does not guarantee complete financial validity or correctness beyond those checks. The benchmark evaluates the full tested configuration, not the model in isolation.

Frequently asked questions (FAQ)