# ClawBench ![Category](https://img.shields.io/badge/Category-Agent-1f6feb) ![Metric](https://img.shields.io/badge/Metric-Success%20Rate-2da44e) ![Saturation](https://img.shields.io/badge/Saturation-Low-d29922) ![Created](https://img.shields.io/badge/Created-2026-555) > **Quick answer:** ClawBench evaluates AI browsing/computer-use agents on everyday tasks performed against real, live, production websites rather than sandboxed clones or offline snapshots — deliberately testing the gap between sandbox performance and the messiness of real-world sites (dynamic layouts, ads, auth flows, rate limits, and unannounced UI changes). ## At a Glance **What it tests:** Whether an agent can complete common online tasks (search, form-filling, checkout-style flows, information lookup) against unmodified, live production websites. **Why it matters:** Many web-agent benchmarks run against static clones or sandboxed replicas of websites, which understate real-world difficulty. ClawBench's live-site design surfaces failure modes — layout drift, anti-bot measures, session/auth handling — that only appear outside a controlled sandbox. **Known limitations:** Because it targets live sites, task difficulty and even task validity can shift over time as target websites change, making long-term score comparability harder than with a static benchmark. Detailed public leaderboard scores were not available at the time of writing. ## What ClawBench Measures ClawBench departs from the sandboxed-website convention used by most browsing-agent benchmarks (which typically run against frozen clones or offline replicas for reproducibility). Instead, it sends agents against real, currently-live production websites to complete everyday tasks — the kind of search, lookup, and form-based interactions a human would routinely perform online. This design choice trades some reproducibility for realism: agents must handle the actual complexity of production web pages, including dynamic content loading, authentication flows, anti-automation defenses, and UI changes that a static sandbox would never expose. The benchmark's premise is that a meaningful gap exists between how agents perform on curated, sandboxed website benchmarks and how they perform "in the wild" — and that gap is exactly what matters for teams deploying browsing agents in production, where the environment is never frozen in time. ## Benchmark Specifications | Field | Value | |---|---| | Task category | Agent (live web browsing / computer use) | | Metric | Task success rate (%) | | Environment | Real, live production websites (not sandboxed clones) | | Saturation | Low | | Created by | Agent benchmarking research community, 2026 | | Paper | Not yet publicly attributed to a single paper | | Dataset | Not public | ## How ClawBench Is Scored Each task is scored as successfully completed or not, based on whether the agent achieves the specified real-world outcome (e.g., locating specific information, completing a form, or navigating to a required page) on the live target site. Because the underlying websites can change between evaluation runs, ClawBench scores are best interpreted as a snapshot of agent robustness at a point in time rather than a permanently fixed leaderboard. ## ClawBench on Benchgen > No Benchgen results yet — [be the first to run ClawBench](/benchmarks/clawbench). ## ClawBench vs Other Benchmarks | Benchmark | What it tests | Saturation | |---|---|---| | **ClawBench** | Everyday tasks on live, unmodified production websites | Low | | [OSWorld-Verified](/benchmarks/osworld-verified) | Computer-use / GUI agent tasks (sandboxed) | Low | | [ClawMark](/benchmarks/clawmark) | Long-horizon, multi-day "coworker" agent tasks | Low | | [MCP Atlas](/benchmarks/mcp-atlas) | MCP tool-use in agentic workflows | Low | ClawBench and OSWorld-Verified both test computer-use/browsing agents, but OSWorld-Verified runs in a controlled, reproducible sandbox while ClawBench deliberately targets live, uncontrolled production sites. Use ClawBench when you need signal on how an agent handles the unpredictability of real websites, not just a controlled replica. ## Run ClawBench on Your Model Benchgen lets teams evaluate their own browsing/computer-use agents against realistic, evolving web environments and track robustness over time as target sites change — rather than relying solely on static sandbox scores that may overstate real-world readiness. ## Related Models & Benchmarks - Related benchmarks: [OSWorld-Verified](/benchmarks/osworld-verified) · [ClawMark](/benchmarks/clawmark) · [MCP Atlas](/benchmarks/mcp-atlas) ## Frequently Asked Questions

What is ClawBench?

ClawBench is a benchmark that evaluates AI browsing/computer-use agents on everyday tasks performed against real, live production websites rather than sandboxed clones, testing robustness to real-world web complexity.

What does a good ClawBench score look like?

Detailed public leaderboard scores were not available at the time of writing; the benchmark is positioned as a robustness check on live sites rather than a fully mature, saturated leaderboard.

Who created ClawBench?

ClawBench comes from the broader 2026 agent-benchmarking research community and is referenced alongside other harness-focused agent benchmarks such as Harness-Bench.

How does ClawBench differ from OSWorld?

OSWorld and OSWorld-Verified run computer-use agents against controlled, reproducible sandbox environments. ClawBench instead runs agents against real, currently-live production websites, deliberately exposing agents to dynamic content, anti-automation defenses, and UI drift that a sandbox would not include.

ClawBench — Evaluations