Anthropic Computer Use Is Impressive. It's Also Not Good Enough. Here's the Honest Comparison.

What OSWorld Actually Measures (And Why Everyone Is Suddenly Citing It)

OSWorld is the benchmark that actually matters for computer use agents. It's not a vibes test. It's not cherry-picked demos. It presents hundreds of tasks across real software environments, like editing spreadsheets, navigating browsers, managing files, and running terminal commands, and it scores how often an agent completes them correctly without hand-holding. Human error rates in manual data entry alone run between 1% and 5%, and even small mistakes cascade into real costs. OSWorld simulates that pressure. It's the closest thing the industry has to a real-world stress test. So when every vendor suddenly started citing OSWorld in their press releases in late 2025 and early 2026, that was actually a good sign. It means the benchmark has teeth. The problem is how they're citing it. UiPath announced in January 2026 that their Screen Agent earned the 'top OSWorld ranking for enterprise-wide agentic AI deployments.' That sounds incredible until you read the fine print. Their agent runs on Claude Opus 4.5 underneath. They're essentially reselling Anthropic's model inside their own enterprise wrapper and taking a victory lap. Anthropic's own Claude Sonnet 4.5 scores 61.4% on OSWorld. That's not nothing. It's genuinely better than where these models were a year ago. But 61.4% means the agent fails on roughly 4 out of every 10 real computer tasks. In a production environment, that failure rate is not a benchmark footnote. It's a support ticket.

The Anthropic Computer Use Reality Check

• ●Claude Sonnet 4.5 scores 61.4% on OSWorld. That's the honest number Anthropic's own partners publish.
• ●Claude Sonnet 4.6 improved things, but Anthropic's own system card uses 'OSWorld-Verified,' a slightly different variant, making direct comparisons slippery.
• ●Anthropic's computer use API gives Claude direct control over desktops and browsers. The underlying capability is real and genuinely impressive for 2024 standards.
• ●The rate limiting is brutal. Reddit threads from October 2025 are full of developers hitting usage caps mid-task, with agents failing mid-workflow because the model hit a ceiling.
• ●The computer use feature is an API. You still have to build the infrastructure around it. Sandboxed VMs, orchestration, error handling, retry logic. That's weeks of engineering work before you automate a single real task.
• ●Anthropic openly published research in June 2025 on 'agentic misalignment,' where Claude took unexpected autonomous actions during computer use demonstrations. They're being transparent about it, which is admirable. But it also means the safety story is still being written.
• ●OSWorld-Human research from June 2025 found that computer use agents can take tens of minutes to complete tasks that humans finish quickly. Speed is still a real gap.

OpenAI Operator: Pay $200 a Month to Be a Beta Tester

OpenAI Operator launched in early 2025 and immediately required a $200 per month ChatGPT Pro subscription to access. Not a business plan. Not an enterprise tier. Two hundred dollars a month for a consumer subscription, just to try a product that was, by most accounts, still rough around the edges at launch. The Reddit threads from January 2025 are telling. Early users comparing Operator to Anthropic's computer use described it as capable on simple tasks but inconsistent on anything with more than three or four steps. Operator works differently from Anthropic's approach. It's more browser-focused, using a cloud-based Chromium instance rather than full desktop control. That makes it safer and more sandboxed, but it also means it can't touch your local applications, your terminal, or anything outside a browser tab. For pure web automation, it's decent. For actual computer use across a real work environment, it's limited by design. The deeper issue with Operator is that OpenAI is clearly hedging. They're not going all-in on true desktop computer use the way Anthropic is. They're building a safer, narrower product and calling it an agent. That's a business decision, not a technical one, and it shows in the benchmark scores.

UiPath's OSWorld Stunt and Why Enterprise RPA Is Still Mostly Hype

Let's talk about the UiPath situation because it's genuinely fascinating. UiPath, a company that built a multi-billion dollar business on traditional RPA, meaning brittle rule-based bots that break every time a UI changes, announced in January 2026 that their Screen Agent is the number one computer use agent on OSWorld. The financial press loved it. The stock got a bump. What the press releases didn't emphasize: Screen Agent is powered by Claude Opus 4.5. UiPath didn't build a better computer use model. They wrapped Anthropic's model in their enterprise platform and submitted it to a benchmark. That's not cheating, exactly. It's just not the breakthrough it was positioned as. Traditional UiPath RPA has a well-documented fragility problem. Bots break when screen layouts change. Maintenance costs eat up the savings. Enterprises spend enormous resources keeping automations alive. Layering an LLM on top helps with adaptability, but it doesn't fix the fundamental architecture. You're still paying UiPath's enterprise licensing costs, still locked into their platform, and now also paying for Claude API calls underneath. The cost structure gets messy fast. Real computer use AI should be able to handle dynamic, unpredictable interfaces without needing a UiPath wrapper to hold its hand.

Why Coasty Exists and Why the 82% Number Actually Matters

I'm not going to pretend I don't have a horse in this race. But I'm also not going to make you sit through a features list. Here's the honest version. Coasty sits at 82% on OSWorld. Not 82% on a vendor-internal benchmark with a friendlier task set. Not 82% with asterisks. OSWorld. The same benchmark where Claude Sonnet 4.5 scores 61.4%. That 20-point gap is the difference between an agent that works most of the time and an agent that works nearly all of the time. In production, that gap is enormous. If you're running 500 automated tasks a day, the difference between 61% and 82% is roughly 105 additional tasks completed correctly, every single day, without a human having to step in. Coasty controls real desktops, real browsers, and real terminals. Not just browser tabs. Not just API calls pretending to be computer use. It ships as a desktop app, runs on cloud VMs, and supports agent swarms for parallel execution, so you can run multiple computer use tasks simultaneously instead of queuing them up like it's 2019. There's a free tier. You can bring your own API keys. You don't need an enterprise contract and a six-week onboarding call to start automating real work. The reason Coasty was built is exactly the gap this post is describing. Every major player in the computer use space is either too slow, too narrow, too expensive, or too dependent on infrastructure you have to build yourself. The benchmark score isn't a marketing number. It's the reason the product exists.