OSWorld 2.0 just dropped and the headline numbers look impressive. GPT-6 Astra leads at 72.6%. That sounds like a win until you actually look at the leaderboard and see how many models are clustered in the 50s, the 40s, even the 30s. This is the benchmark that's supposed to prove AI agents can actually use a computer, yet it exposes exactly how far we are from reliable computer use in production.
The OSWorld 2.0 leaderboard tells an uncomfortable story
The September 2026 OSWorld 2.0 leaderboard shows GPT-6 Astra at 72.6% with 22 models competing. Most of the others are in the 50s or lower. This isn't a tight race. It's a canyon. Even the models that claim to be cutting edge are struggling with basic desktop tasks like managing windows, reading menus, and following multi-step workflows. The gap between the top tier and the rest is so large it's almost embarrassing. Stanford's 2026 AI Index Report notes OSWorld performance rose from 12% to 66.3% over the past year, which sounds like progress until you realize we're still nowhere close to human capability for complex, open-ended tasks.
Human experts are still miles ahead
- Human baseline on OSWorld is around 72.4% according to OpenAI's own docs
- Top frontier models like GPT-5.4 and Opus 4.6 barely edge past human performance
- Most agents fail on simple things like reading error messages or finding the right menu
- The 2026 AI Index Report confirms models now meet or exceed human baselines on some benchmarks, but OSWorld is still waking up from the 12% dark ages of 2025
The uncomfortable truth: OSWorld 2.0 was designed to expose exactly this gap, and on September 5, 2026, the leaderboard tells a genuinely uncomfortable story about how far most AI agents still are from reliable computer use.
Why OSWorld actually matters for real work
OSWorld isn't some abstract academic benchmark. It tests agents on real desktop tasks across operating systems. That means opening apps, moving windows, clicking buttons, typing in forms, and handling errors when things go wrong. These are the exact skills a computer use agent needs to automate actual work. But the results show most agents can't consistently do the basics without constant human supervision. You can't trust an agent that gets stuck on a dialog box or hallucinates a menu it can't see. That's why failure modes matter more than any benchmark score.
Why your competitor's claims about computer use are probably lies
- Most vendors cherry-pick easy tasks and hide their failure rates
- Relying on mock environments instead of real OSWorld verification
- Claiming 'near human' performance without independent verification on the official OSWorld leaderboard
- Ignoring that agents break when they encounter unexpected error messages or UI changes
Why Coasty Exists
This is where Coasty comes in. We don't just claim better performance. We independently verified 82.81% on the official OSWorld Verified leaderboard at osworld-v1.xlang.ai. That's not a typo. Our Coasty computer use agent is near the top of the leaderboard, with our own in-house model hitting 85.6% on public OSWorld results. We control real desktops, browsers, and terminals. Not just API calls. Our agent can swarm across multiple machines to run parallel workflows, and we support both desktop apps and cloud VMs. We also support BYOK so you can bring your own keys if that matters to you. There's a free tier if you want to try it yourself without committing. When you compare to manual work or other agents, the gap becomes obvious.
OSWorld 2.0 shows progress, but let's not pretend we've won. The leaderboard is full of models that can't reliably use a computer. If you're still paying humans to copy-paste data or click through the same menus day after day, you're doing it wrong. Coasty is the #1 computer use agent for a reason. Check it out at coasty.ai and stop letting your team waste time on work an AI agent can already do better.
Want to see this in action?
View Case Studies