Benchmarks

Standardized evaluations of LLM reliability on real-world tasks. Open methodology, reproducible results, real tool calls against real APIs.

Language Active

Thai Speech-to-Text

The models best at Thai drop every English word

Privacy Active

PII Detection

Flagship-quality PII detection from a 3× smaller model

Business Tools Active

Sheets Benchmark

Tool architecture matters more than model choice

Business Tools In progress

Email Benchmark

Draft quality, thread handling, attachment accuracy

Business Tools In progress

Calendar Benchmark

Scheduling accuracy, conflict detection, timezone handling

Developer Tools In progress

GitHub Benchmark

PR reviews, issue triage, code search accuracy

Computer Use In progress

Playwright Benchmark

Web task completion, UI resilience, action policy safety

Robotics In progress

Robotics VLA Benchmark

Vision-language-action task completion in simulated environments

Data Science In progress

Data Analysis Benchmark

SQL generation, data cleaning, visualization accuracy

3D / Spatial In progress

OpenUSD Benchmark

Scene graph edits, material assignments, transform accuracy

3D / Spatial In progress

STL Benchmark

Mesh manipulation, geometric accuracy, printability validation

Why we benchmark

Most LLM benchmarks test knowledge or reasoning in isolation. We test what matters for production: can a model reliably complete multi-step tasks using real tools against real APIs, without corrupting your data?

Every benchmark uses Daslab's tool-calling infrastructure with real Google, GitHub, and Slack APIs. Tasks are defined as expected input/output pairs with cell-level validation. Results are fully reproducible.

Real APIs

Not simulated. Every tool call hits the actual Google Sheets, Gmail, or GitHub API. Errors are real errors.

Cell-level validation

Success isn't "did it respond" — it's "does the spreadsheet contain exactly the right values in the right cells."

Open methodology

Task definitions, evaluation scripts, and raw results are all public. Run the same benchmark on your own infrastructure.

Cost-aware

We track tokens, retries, and dollar cost per task. The cheapest model that works is often the right choice.