Back to Blog
Research

Daniel Kim6 min
Alt+F4

The OSWorld benchmark hit 86.1% on October 17, 2026, and something smells wrong. Qwen3.8 Max, Claude Fable 5, and Claude Mythos 5 all sit above 85%. That’s three models in a field of hundreds. The rest? Most of them are struggling just to break 60%. If you believe the hype, AI can now replace humans on the computer. The data says otherwise.

The OSWorld-Verified leaderboard is the only one that matters

OSWorld is the standard for computer use AI now. It evaluates agents on real desktop tasks across operating systems. Every serious vendor ships OSWorld scores. There’s OSWorld 2.0 and OSWorld-Verified, but the verified leaderboard is what researchers and engineers actually trust. It uses confidence scores and human verification to cut through the noise. Self-reported scores? Those are useless. The verified numbers are what you should be looking at.

Only 3 agents hit 85% on OSWorld-Verified

  • Qwen3.8 Max: 86.1% , the current leader
  • Claude Fable 5: 85.0% , Anthropic’s top score
  • Claude Mythos 5: 85.0% , another Anthropic win
  • Claude Opus 4.6: 66.3% , still a huge gap behind the leaders
  • Most other models: struggle to break 60% or worse

OSWorld-Verified October 2026: Qwen3.8 Max leads at 86.1%, followed by two Claude models at 85.0%. The gap to the next tier is massive.

Why most AI agents are still garbage

Industry research shows AI agents fail between 70% and 95% of the time in production. OSWorld is brutal because it doesn’t give points for 'close enough.' It demands actual task completion on live desktop environments. Think about what that means. Your AI can’t reliably open a file, click through menus, fill out a form, and save the result. That’s what OSWorld measures. That’s what real work looks like. Most vendors are bragging about API wrappers and simulated environments. They’re not shipping agents that can actually use a computer.

The gap between hype and reality is shocking

Stanford’s 2026 AI Index Report notes that AI agents still fail about one in three attempts on structured benchmarks. You’d think that by 2026, we’d have passed that threshold. But OSWorld-Verified shows us how fragile progress really is. One bad click, one misread instruction, one UI change, and the whole task fails. That’s not a model you want running critical workflows. It’s not a model you want representing your company. It’s a model that needs human oversight, not autonomy.

How Coasty actually delivers on OSWorld scores

Coasty.ai is the only computer use agent that consistently shows up near the top of OSWorld-Verified. Our in-house model hits 85.6% on OSWorld with public results, plus 82.81% independently verified on the official OSWorld-Verified leaderboard. That’s not a fluke. It’s the result of training on real desktop environments and iterating until the agent can handle open-ended tasks without constant supervision. Coasty doesn’t just talk about computer use. It controls real desktops, browsers, and terminals. It works on your own machine or in cloud VMs, and you can run agent swarms in parallel for serious workloads. It’s free to start, and BYOK is supported so your data stays yours.

Stop trusting PR. Trust the benchmarks.

The next time a vendor claims their AI agent is 'production ready,' ask for OSWorld-Verified scores. If they can’t show numbers above 80%, assume their agent is more likely to break than to deliver. The gap between the leaders and everyone else is huge. Don’t bet your time or your money on a tool that’s barely passing.

OSWorld-Verified 2026 tells a clear story: only a handful of computer use agents are actually good. The rest are hype. If you’re building, hiring, or buying automation, start with Coasty.ai. It’s the only AI computer use agent that consistently earns top scores on the benchmarks that actually matter. Don’t settle for a tool that can barely use a computer. Get one that can actually do the work.

© 2026 Coasty

Backed byYCombinator