Benchmarks
Agentic accuracy and token efficiency on ShortcutBench, plus engineering targets for Mog's Rust/WebAssembly compute core.
Agentic Benchmarks
Mog is designed to be operated by AI agents. We measure API quality using ShortcutBench (v25), a benchmark suite that evaluates how accurately and efficiently LLM agents complete spreadsheet tasks through each SDK.
Accuracy on ShortcutBench (v25)
End-to-end accuracy of AI agents completing spreadsheet tasks via each SDK's API surface.
ShortcutBench v25 measures how accurately an LLM agent can complete common spreadsheet operations (formatting, formulas, charts, pivot tables) through each SDK. Higher is better.
Token Efficiency vs Competitors
Average tokens required to complete the same spreadsheet task, normalized to Mog (lower is better for the SDK).
Mog's API requires fewer tokens per task on average. SpreadJS requires ~2.1x the tokens and OfficeJS ~1.9x to accomplish the same operations.
Performance Targets
Engineering targets based on Rust/WASM architecture. Reproducible benchmark harnesses will be published alongside the open-source release.
Recalculation (10K formulas)
Full dependency-graph recalculation of 10,000 formulas
Target based on Rust/WASM compute architecture. Competitor comparisons will be added after independent benchmarking.
XLSX parse (10MB file)
Time to parse and render a 10MB .xlsx file
Target based on Rust XLSX parser compiled to WASM. Actual numbers will vary by file complexity.
Cold start (WASM load)
Time from page load to interactive spreadsheet
Target for initial WASM download, compile, and first render on a median connection.
Methodology
The performance targets above are based on the architectural advantages of a Rust/WASM compute engine over JavaScript-based alternatives. Reproducible benchmark harnesses, test data, and full methodology will be published alongside the open-source release. Until then, treat these numbers as engineering goals, not verified claims.
Rust compute, everywhere
Mog's compute core is written in Rust and compiled to WebAssembly for browsers or native bindings for Node.js — the same engine on every platform. We expect formula evaluation, recalculation, and XLSX parsing to run at near-native speed with no server round-trips.