# Claw-Eval ![Category](https://img.shields.io/badge/Category-Agent-1f6feb) ![Metric](https://img.shields.io/badge/Metric-Pass%5E3-2da44e) ![Tasks](https://img.shields.io/badge/Tasks-300-8957e5) ![Saturation](https://img.shields.io/badge/Saturation-Low-d29922) ![Created](https://img.shields.io/badge/Created-2026-555) > **Quick answer:** Claw-Eval is a 300-task, 9-category agent benchmark that has become a cross-lab reference point — it's referenced or used internally by Meta, Kimi, Qwen, and Tencent as a "trustworthy" evaluation baseline. It uses Pass^3 scoring (each task is attempted 3 independent times and judged on consistency across trials) specifically to reduce the noise that single-run pass/fail evaluation introduces into agent benchmarking. ## At a Glance **What it tests:** General agent task-completion ability across 300 human-verified tasks spanning 9 categories, scored with a 3-trial consistency methodology rather than a single pass/fail attempt. **Why it matters:** Single-run agent evaluation is notoriously noisy — the same agent can pass or fail an identical task on different attempts due to sampling variance. Claw-Eval's Pass^3 methodology directly addresses this, which is likely why it has been adopted as a reference/sanity-check benchmark across multiple frontier labs. **Known limitations:** No single originating paper with full authorship has been consistently cited alongside Claw-Eval; it functions more as a shared cross-lab reference set than a benchmark with one canonical public leaderboard. ## What Claw-Eval Measures Claw-Eval evaluates general agent capability across 300 human-verified tasks organized into 9 categories, designed to be broad enough to serve as a cross-lab sanity check rather than a narrow, single-domain test. Its defining methodological choice is Pass^3 scoring: rather than running each task once and recording pass/fail, Claw-Eval runs each task 3 independent times and incorporates trial-to-trial consistency into the score. This directly targets a known weakness of single-run agent benchmarks — an agent that passes a task once by chance is scored differently from one that reliably passes it every time. Because of this trust-oriented design, Claw-Eval is cited as a reference benchmark by multiple frontier labs (Meta, Kimi, Qwen, and Tencent), typically as one data point among several rather than the sole basis for capability claims — a role similar to how AlpacaEval or MT-Bench function as widely trusted secondary checks alongside a lab's primary benchmark suite. ## Benchmark Specifications | Field | Value | |---|---| | Task category | Agent (general capability, cross-lab reference) | | Metric | Pass^3 (pass rate across 3 independent trials) | | Number of tasks | 300, human-verified | | Categories | 9 | | Adopted/cited by | Meta, Kimi, Qwen, Tencent | | Saturation | Low | | Created by | Cross-lab agent evaluation reference | | Paper | Not consistently attributed to a single public paper | | Dataset | Not public | ## How Claw-Eval Is Scored Each of the 300 tasks is attempted 3 independent times per agent under test. The Pass^3 metric credits an agent based on how consistently it succeeds across those 3 trials, rather than treating a single successful attempt as full credit. This makes Claw-Eval more resistant to sampling-variance noise than one-shot pass/fail benchmarks, at the cost of requiring 3× the inference compute per task to produce a score. ## Claw-Eval on Benchgen > No Benchgen results yet — [be the first to run Claw-Eval](/benchmarks/claw-eval). ## Claw-Eval vs Other Benchmarks | Benchmark | What it tests | Tasks | Saturation | |---|---|---|---| | **Claw-Eval** | General agent capability, 3-trial consistency scoring | 300 | Low | | [Harness-Bench](/benchmarks/harness-bench) | Harness quality, holding model & task fixed | 106 | Low | | [MT-Bench](/benchmarks/mt-bench) | Multi-turn conversational quality | — | High | | [Arena-Hard v2](/benchmarks/arena-hard-v2) | Preference-based model comparison | — | Medium | Claw-Eval plays a similar cross-lab "trust check" role to MT-Bench and Arena-Hard, but applied to agent task completion rather than conversational quality, and with a multi-trial (Pass^3) methodology specifically built to reduce single-run noise. ## Run Claw-Eval on Your Model Benchgen lets teams run multi-trial (Pass^k) agent evaluations against their own models and harnesses, applying the same consistency-scoring approach Claw-Eval uses to distinguish reliable task completion from lucky single-run passes. ## Related Models & Benchmarks - Related benchmarks: [Harness-Bench](/benchmarks/harness-bench) · [MT-Bench](/benchmarks/mt-bench) · [Arena-Hard v2](/benchmarks/arena-hard-v2) ## Frequently Asked Questions

What is Claw-Eval?

Claw-Eval is a 300-task, 9-category agent benchmark that uses Pass^3 (3-trial) scoring to reduce single-run noise, and is referenced as a cross-lab evaluation baseline by Meta, Kimi, Qwen, and Tencent.

What does Pass^3 scoring mean?

Each task is attempted 3 independent times, and the agent's score reflects how consistently it succeeds across all 3 trials rather than crediting a single successful attempt — reducing the effect of sampling variance on the final score.

Who created Claw-Eval?

Claw-Eval does not have one consistently cited originating paper; it functions as a shared cross-lab reference benchmark used internally by multiple frontier labs.

Is Claw-Eval saturated?

Public score data is limited, but its 9-category, 300-task, 3-trial design is intended to remain a meaningful differentiator across agent capability levels rather than a fully solved benchmark.

Claw-Eval — Results

Claw-Eval — Evaluations