The easiest way to
evaluate any model.

Run frontier agentic evals on any model, including your own. Launch them from your agent, and see exactly where each model fails. No infrastructure to build.

Browse 274 benchmarks →

Claude Code · via MCP

run_benchmark("terminal_bench_2_1", models=3)

Your model

your-model-v3 · endpoint connected

Ready to run

Terminal-Bench 2.1 · from $0.04 per task

Failure analysis

An AI-written report on where each model goes wrong, with the failing steps flagged

Illustrative

Backed by

Working with

Frontier labs
Enterprise AI teams
Vertical AI companies
Investment funds
Research groups

Frontier benchmarks

Run frontier agentic evals

The benchmarks in every frontier model card, ready to run on any model in one click: agentic coding, terminal work, computer use, tool use and hard reasoning. 274 in the catalogue, 107 of them agentic.

Compare your model on the benchmarks that matter, not on one nobody else uses.

Browse the catalogue →

In a frontier model cardprice per task, from

SWE-bench Verified

OpenAI

$0.03

SWE-Bench Pro

Scale AI

$0.27

Terminal-Bench 2.1

Terminal-Bench

$0.04

OSWorld

XLANG Lab

$0.11

tau-bench

Sierra

$1.99

Humanity's Last Exam (text)

CAIS & Scale AI

$0.01

Agents

Run evals directly from your agent

Add our MCP server to Claude Code, Cursor or Codex and ask in plain English. Your agent searches the catalogue, shows you the estimated cost, launches the run and reads back the results.

“Run SWE-bench Verified on these three models and tell me where they differ.”

Run in your terminal, then sign in when Claude Code asks.

Read the MCP docs: sign-in, examples and every tool →

Your model

Evaluate your own model

Point us at any OpenAI-compatible endpoint (vLLM, TGI, a cloud deployment, a private gateway). We check the connection, then score it on the same benchmarks as the frontier models, side by side.

Saved once, connection-checked before anything runs
Your weights stay behind your endpoint
Private benchmarks visible only to your organisation

Results chart on Terminal-Bench 2.1: your-model-v3 scores 78%, ahead of an Anthropic model at 74% and an OpenAI model at 72% — Example run

Agentic, safely

Run agentic benchmarks safely, without the infrastructure hassle

Agentic benchmarks need a sandbox per task, the right environment image, an agent harness and graders. We run all of it, and keep each agent contained: every task gets its own isolated sandbox, and every finished trajectory is screened for cheating (looking up the answer, tampering with the grader) and misuse (trying to escape the sandbox, reach credentials or attack other systems). Run limits stop runaway spend.

No Docker images to build, no sandboxes to provision, no grader prompts to maintain.

See the agentic benchmarks →

Understand the failures

The fastest way to improve a model is to see exactly where it fails

A score tells you who lost. The trial reader shows you why: every step each agent took, with the steps that errored flagged so you can jump straight to them.

Read every trajectory, with screenshots for computer-use runs
Compare models on the same task, side by side
Share a no-login link, or download every trajectory

The trial reader on a failed Terminal-Bench task: 7 steps, 2 flagged, with the failing alembic upgrade step expanded — Example run

From question to answer in four steps

The same flow whether you use the web app or ask your agent.

Pick a benchmark

Search 274 benchmarks, or bring your own dataset.

See the cost

Choose models and get an estimated cost range before anything runs.

Run in sandboxes

Every task in its own environment, graded automatically.

See where it fails

Scores, cost and the steps that went wrong, per model.

Find out where your model fails.

We're in invite-only beta. Request an invite and we'll get you set up.

Already invited? Log in

The easiest way to evaluate any model.