The whole setup is built around authoring benchmarks and throwing models at them to see who survives. It has native support for the harbor authoring system, and if you're feeling fancy, you can use the Vercel AI SDK for custom tool authoring. They've already integrated TerminalBench 2.0 to prove their harbor task orchestrator actually works, though that part requires a paid account because provisioning sandboxes isn't free (shocker).
If you don't feel like paying for a sandbox yet, there are several public demos where you can pick a model from a list and watch it struggle—or succeed—in real-time. I've seen a few of these available:
GSM8K: For when you want to see if the model can actually do basic math without losing its mind.ARC-Challenge: Testing if the LLM has any actual reasoning capabilities or is just a glorified autocomplete.TruthfulQA: A great way to see which models are confident liars.Medical AI Failure Atlas: Because trusting an AI with your health is a bold choice, so you might as well see where it fails first.SkateBench: Because why not?
For anyone looking for a practical tutorial on how to actually validate a model's performance without building a custom pipeline from scratch, this is a decent shortcut. You just pick the model, pick the scenario, and let the system do the systematic testing. It’s a much cleaner AI workflow than the usual "I tried three prompts and it seemed to work" method of benchmarking.
The tool is still evolving, so there's plenty of room for it to get better. The documentation could probably be tighter, and the spend limits/configuration side of things always feels a bit clunky in these types of platforms. But as far as a deep dive into model capabilities goes, having a browser-based interface is a massive quality-of-life upgrade over staring at a JSON file of results.
https://trunchbull.dev/sandboxes
Next Gemini hit 1 billion users faster than any other Google product →