trunchbull
daytona startup grid · november 2026
The place to discover, run
and share benchmarks.
Find the right benchmark, run it on any model with a clean developer experience, private evaluation cases and real environments, then share the results.
the problem
Choosing a model is still guesswork
Teams switch models every few weeks on the strength of leaderboards and gut feel, because testing on their own tasks is too much work.
01
Benchmarks are scattered
They live across papers, repos and leaderboards, each with its own harness. Finding one and rerunning it is a project.
02
Sharing leaks the answers
Once test cases are public they end up in training data, and the scores stop meaning much.
03
Agents need somewhere to run
Agent benchmarks need a clean, isolated machine for every attempt. Most teams don’t have one.
trunchbull
One place to find, build, run and share benchmarks
discovery
A public registry
Browse benchmarks, their public cases and scoreboards, and run them on your own models.
developer experience
A benchmark is a repository
Plain files, no SDK. Push, and Trunchbull builds a release you can run on any model.
private cases
Share scores, not answers
Cases are private by default. They count toward public scores but are never shown.
environments
A fresh machine per attempt
Each attempt gets its own sandbox with the case’s image and network rules.
discovery
A registry for finding benchmarks
30
public benchmarks
204
models on scoreboards
8
official, run in one click
2,728
public cases to inspect
Every listing shows its public cases, tools and grader. Scoreboards only include runs their owners chose to share.
developer experience
A benchmark is just a repository
No SDK to install
Evals and tools are plain TypeScript objects. Trunchbull handles bundling and validation.
Every push is a release
Each push compiles an immutable release, so a score always points to exact cases.
Run any model
Pick models from the catalog, or bring your own OpenRouter key.
private evaluation cases
Publish the score, keep the answers
Private by default
Prompts, expected answers and graders stay hidden from everyone but the owner.
Still counted
Private cases are part of every public score, so results can’t be gamed by studying them.
Public samples when you want them
Choose a few cases to show in full, with transcripts, so readers can see what’s tested.
example public result · 1,200 cases
Scores include every case. Only the public samples can be opened.
environments
Every attempt gets its own environment
Start from the case’s pinned container image
Apply the network rules the author set for that case
Run the agent, then check its work with the case’s verifier
Destroy the sandbox. A retry starts from scratch.
| network rule | what the agent can reach |
|---|---|
| None | Nothing, including DNS |
| Allowlist | Only the listed domains and IP ranges |
| Open | The public internet |
Managed sandboxes by default, or your own provider through a self-hosted runner: Daytona, Modal, E2B, Fly or Docker.
sharing
Every benchmark and run gets a public page
Anyone can read the scores, then run the same benchmark on their own models. These official benchmarks are ready to run today.
why daytona
Daytona is already one of our sandbox providers.
Teams can run Trunchbull trials on their own Daytona account with our published CLI. Their credentials stay on their machine, and every attempt still gets a new sandbox that’s destroyed at the end.
trunchbull runner --driver daytona
business model
Free to start, then a monthly plan
| plan | price / month | run cases / month | concurrent sandboxes |
|---|---|---|---|
| Free | $0 | 1,000 | 1 |
| Team | $20 | 20,000 | 5 |
| Growth | $50 | 50,000 | 10 |
| Scale | $100 | 100,000 | 20 |
| Enterprise | $250 | 250,000 | 50 |
Usage beyond the plan is paid from prepaid credits
Bring your own model key on every plan, Free included
Self-hosted runners for teams with their own compute
the ask
What we’d do with Daytona credits
01
Finish the Daytona runner
Our Modal and Docker runners are tested live. The Daytona runner is written and next.
02
Enforce each case’s network rules
Map none, allowlist and open onto Daytona’s per-sandbox network settings.
03
Run our catalog on Daytona
Run the official benchmarks end to end so teams can rely on it.
The ask: Startup Grid credits, plus Tier 3 or a custom tier so our runner can use Daytona’s per-sandbox network settings.
trunchbull
Thank you.
Browse the registry at trunchbull.dev/registry, or start free at trunchbull.dev.