Free
$0
forever
- 1,000 tool calls per month
- Up to 5 published benchmarks
- Up to 5 published tools
- No sandbox provisioning
A Tool-Native Benchmarking Platform
Run any model against any Benchmark
Deploy tools from GitHub, run them against any OpenRouter model, and inspect every decision under controlled budgets.
BENCHMARK READOUT / RUN 08422026-07-19 09:41:02 UTC LIVE
ModelScoreDeltaState
Claude Opus 494.2base
GPT-591.8-2.4%
Gemini 2.5 Pro88.5-5.7%
2 REGRESSIONS DETECTED14 TOOLS / $1.84
REMOTE TOOLS
Isolated Workers
MODELS
OpenRouter / BYOK
LIMITS
Steps · tokens · cost
EVIDENCE
Full run traces
01
Tools as infrastructure
Turn a public repository into a model-ready tool.
Give us a GitHub URL. Trunchbull validates the contract, builds the repository, and deploys it into an isolated Cloudflare Worker—ready for any benchmark configuration.
- Vercel AI SDK-compatible schema
- Versioned, reproducible deployments
- Remote execution with hard timeouts
TOOL DEPLOYMENTREADY
Repository URL
github.com/acme/research-toolsmain
Contract validated3 tools
Worker bundle built184 KB
Deploying to edge34 regions
03
TOOLS
34
REGIONS
ISO
RUNTIME
02
Isolated execution
Give agents a real machine without giving them yours.
Run terminal and code-execution evaluations inside controlled, image-backed environments. Trunchbull exposes files, commands, processes, and interactive terminals while keeping each workload inside an isolated runtime.
- Container isolation with explicit resource limits
- File, command, process, and terminal tools
- Image-backed environments for repeatable runs
SANDBOX ENVIRONMENT / LEASE 7F2AISOLATED
terminal-agent / ubuntu:24.04
IMAGE PINNED · PRIVATE RUNTIME
file_readREADY
command_executeREADY
process_listREADY
terminal_openREADY
POLICY
NETWORK
LIMITED
COMPUTE
SCOPED
ACCESS
03
Run control
Give every model a hard ceiling.
Define the maximum steps, token usage, and dollar spend before a run begins. Limits are enforced across model responses and tool calls—not checked after the damage is done.
- BYOK through OpenRouter
- Per-model pricing and token accounting
- Deterministic stop reasons
RUN CONTROL / ACTIVEWITHIN LIMITS
CUMULATIVE USAGE / LIVE
6 of 12 steps complete
TOKENSSPEND
HARD LIMIT
00:00NOW / 14:32
PROJECTED SPEND
$3.72
HEADROOM
25.6%
STOP POLICY
HARD
STEPS
06 / 12
TOKENS
18.4K / 40K
SPEND
$1.84 / $5.00
01MODELPlanning tool sequence1.2s
02TOOLsearch_issues({ query: "timeout" })842ms
03TOOLread_file({ path: "src/client.ts" })186ms
04MODELSynthesizing final answer2.4s
ELAPSED 00:14:32STREAMING
04
Deep evidence
See exactly why one model wins.
A score tells you what happened. Traces show you why. Compare tool selection, arguments, latency, errors, token usage, and final outcomes step by step.
- Side-by-side model traces
- Tool-call and argument inspection
- Regression-ready run history
MODEL COMPARISONCOMPLETE
METRICOPUS 4GPT-5
Task success94%87%
Tool accuracy91%72%
Median latency6482
Usage by model
SHARE OF TOTAL RUN COST
$4.82
Claude Opus 4$1.84
GPT-5$1.46
Gemini 2.5 Pro$0.98
Qwen 3 Coder$0.54
AUTO-CYCLING / 3.8S
FROM REPO TO RESULT
Your first useful result in three moves.
01
Connect a tool repository
Paste a public GitHub URL. We validate, build, and deploy it to an isolated Worker.
02
Configure the run
Choose models, prompts, tools, and hard limits for steps, tokens, and spend.
03
Inspect the evidence
Compare outcomes and replay every model response, tool call, error, and cost.
OPEN BETA PRICING
Start free. Upgrade when your runs need more room.
Publish and evaluate at no cost, then unlock managed sandboxes and higher monthly capacity with one fixed plan.
Open Beta
$59.99
per month
- 250,000 tool calls per month
- 16 sandbox hours per month
- Standard-4 sandbox default
- Unlimited published benchmarks
- Unlimited published tools
Open Beta is a recurring monthly subscription that renews until canceled. See our Terms, Privacy Notice, and Refund Policy.
READY WHEN YOU ARE
Stop guessing which model is better.
Bring a tool, choose your models, and leave with evidence you can act on.