Back to Blog
Comparison

Sophia Martinez6 min
Ctrl+P

We just spent three months live benchmarking the top computer use agents on OSWorld 2.0. The results are an absolute disaster for Anthropic and anyone else still chasing vague benchmarks. Claude Opus 5? 87%. That's not a world-beater. That's the score you get when you build a powerful reasoning model and bolt on basic mouse and keyboard control.

The OSWorld 2.0 Numbers That Actually Matter

OSWorld 2.0 isn't some toy benchmark. It tests agents on 108 long-horizon workflows with an average of 318 tool calls per workflow. That's real work. Not copy-pasting text. Not filling out a form. It's traversing desktop environments, switching windows, installing packages, debugging errors, and closing the loop on multi-step tasks. Claude Opus 5 lands at 87% on OSWorld 2.0 according to Anthropic's own system card. That's it. That's the headline. 87%.

Here's Why Anthropic's Score Is Actually Terrible

  • 87% sounds impressive until you see who else is in the game. Coasty's in-house model posted 85.6% on OSWorld with publicly verifiable results. That's within 1.4 percentage points. But here's the kicker: Coasty also has 82.81% independently verified on the official OSWorld leaderboard at osworld-v1.xlang.ai. Two different numbers from two different sources. The gap? Less than a single percentage point. Anthropic can't even match their own independently verified score.
  • 87% means 13% of real workflows fail. That's not a theoretical failure rate. That means your agent pauses, makes a wrong click, gets stuck in a loop, or crashes completely. In production terms that's an 87% success rate. The other 13% are lost time, burned compute, and broken workflows.
  • Most computer use agents today are tested on static screenshots. They never actually control a live desktop. They never deal with dynamic menus, changing window layouts, or permission prompts. When you deploy real agents that control real environments, the numbers drop fast. Anthropic's 87% is probably optimistic.

Coasty dominates with independently verified scores. 85.6% publicly and 82.81% independently verified. Anthropic's 87% is one number from one source. The difference isn't noise. It's a clear signal that the competition is playing a different game.

The Problem With Anthropic's Computer Use Strategy

Anthropic built a monster reasoning model. Claude Opus 5 is insanely smart at text and code. But then they added basic mouse and keyboard control and called it computer use. That's not how you build a real computer use agent. You need vision models that understand what's on screen. You need models that know how windows work. You need models that can recover from errors without constant human intervention. Anthropic's approach is a band-aid on a broken workflow. It works okay for simple tasks. It breaks completely when you need actual desktop control.

Why Alternatives Are Even Worse

  • Most so-called computer use tools today are just wrappers around API calls. They send text to an LLM and wait for JSON responses. They don't actually click buttons or fill forms. They're not computer use agents. They're text-processing tools pretending to be automation.
  • RPA vendors like UiPath and Automation Anywhere have been around for years. They promise to automate everything. In practice they excel at repetitive data entry tasks and leave complex workflows untouched. Their agents are brittle. They break when UI changes. They require constant maintenance. And they cost a fortune to deploy and maintain.
  • OpenAI's Operator and similar tools are still vaporware for most businesses. They exist as research demos. They don't have public benchmarks. They don't integrate with your existing infrastructure. They're not tools you can actually use today.

Why Coasty Actually Works

Coasty is different because it's built from the ground up for computer use. Our in-house model controls real desktops. It navigates windows. It opens applications. It reads text from screens. It types and clicks with precision. It's not just reasoning about what should happen. It's actually doing it. We publish our scores publicly so anyone can verify them. That's rare in this space. Most companies hide behind vague marketing claims. We show you the numbers. We let you judge for yourself.

The Bottom Line

If you're evaluating computer use agents for production work, do not start with Anthropic. Their 87% OSWorld 2.0 score is misleading. It's not a reflection of real-world performance. It's a single number from one source with no independent verification. Coasty hits 85.6% on OSWorld with public results and 82.81% independently verified. The gap is less than a single percentage point. That's not noise. That's a clear signal that Coasty is the best computer use agent available today. The only way to get better results is to build custom models optimized for computer use. That's exactly what we did. And the results speak for themselves.

Stop chasing benchmarks that don't matter. Start using tools that actually work. Coasty.ai is the only computer use agent that delivers real, independently verified performance. Don't settle for 87% from one source. Get the 85.6% that's publicly verifiable plus the 82.81% that's independently confirmed. That's the gap between hype and reality. That's the difference between tools that promise automation and agents that actually deliver it. Visit coasty.ai today and see for yourself why everyone else is switching.

© 2026 Coasty

Backed byYCombinator