Datasets:
license: mit
language:
- en
pretty_name: llm-arithmetic
size_categories:
- n<1K
task_categories:
- question-answering
tags:
- llm-evaluation
- arithmetic
- finance
- claude
configs:
- config_name: rows
data_files:
- split: train
path: results.json
field: rows
llm-arithmetic
Author: Dmitriy Semenkevich (dimhold)
This dataset measures whether 4 Claude models can do 10 pieces of SaaS arithmetic (MRR, ending cash, LTV, runway) with every tool taken away.
Headline, from the repository: exact answers were 10 / 10 for claude-opus-5, 9 / 10 for claude-sonnet-5, 3 / 10 for claude-haiku-4-5 and 9 / 10 for claude-fable-5; all 4 models were within 1% on 10 / 10 (RESULTS.md).
This is a mirror. The canonical home is the Zenodo record and the GitHub repository; criteria, limitations and full tables live there.
- Paper: https://dimhold.by/papers/llm-arithmetic/
- Code and full write up: https://github.com/dimhold/llm-arithmetic
- DOI: https://doi.org/10.5281/zenodo.22128841
Files
| path | what it is |
|---|---|
results.json |
every call: tasks with expected values computed in code, and one row per model and task under rows |
transcript.md |
every question and every full reply |
RESULTS.md |
the result tables |
How it was produced
Run on 2026-08-16 through the Claude Code CLI with built in tools and MCP servers disabled. Tasks are generated from seed 20260812, ground truth is computed in run.ts and never by a model, and the answering model is verified against modelUsage. The scripts that turn these files into every number are in the GitHub repository. One local path in results.json is written as <private-project>. Nothing else was edited.
Licence
MIT, the licence of the repository, covering code and data.
Citation
@misc{semenkevich2026llmarithmetic,
author = {Semenkevich, Dmitriy},
title = {llm-arithmetic: can Anthropic's models do SaaS arithmetic without a calculator?},
year = {2026},
doi = {10.5281/zenodo.22128841},
url = {https://doi.org/10.5281/zenodo.22128841}
}