The 2026 AI Agent Benchmark Results Are Out and Most 'Computer Use' Tools Are Still Embarrassingly Bad

What OSWorld Actually Measures (And Why It's the Only Score That Matters)

OSWorld is 369 real computer tasks. Not multiple choice questions. Not 'predict the next token.' Actual desktop work: navigating file systems, filling out forms, using real apps like LibreOffice and Chrome and terminals, handling pop-ups, recovering from errors. It's the closest thing we have to a standardized test for genuine computer use ability. Most other benchmarks are either too narrow (coding only, web browsing only) or too easy to game with clever prompting tricks. OSWorld is hard to fake because it requires an AI agent to actually operate a computer the way a person does. That's why the scores are so much lower than, say, MMLU or HumanEval, where models routinely score in the high 90s. Real computer use is hard. The gap between 75% and 82% on this benchmark represents a genuinely different class of reliability. At 75%, roughly 1 in 4 tasks fails. At 82%, it's fewer than 1 in 5. When you're running hundreds of tasks a day in production, that difference is the difference between a useful tool and a tool you're constantly babysitting.

The 2026 Leaderboard, Ranked Honestly

• ●Coasty: 82% on OSWorld. The actual top score. Runs on real desktops, cloud VMs, and supports agent swarms for parallel execution.
• ●GPT-5.4 (OpenAI): 75% on OSWorld. Just crossed human expert level (72.4%). The internet's current darling.
• ●Claude Opus 4.6 (Anthropic): 72.7% on OSWorld verified. Solid but still technically below the human expert baseline.
• ●Claude Sonnet 4.6 (Anthropic): Improving fast, but Anthropic's own charts show it trailing Opus significantly on computer use tasks.
• ●OpenAI Operator / CUA (older): Built on GPT-4o vision plus reinforcement learning. Benchmarks from early 2025 put it well below 60% on OSWorld-style tasks. GPT-5.4 is a massive jump, but Operator in production still has real reliability complaints.
• ●UiPath and legacy RPA: Not even playing this game. Rule-based automation breaks the moment a UI changes. It's not a computer use agent, it's a very expensive macro recorder.

The Benchmark Controversy Nobody Wants to Touch

Here's where it gets spicy. There's a growing and very loud argument in AI circles that benchmark scores are, to put it politely, not the whole story. A recent piece on Towards AI called LLM benchmarks 'junk science,' pointing out that only 16% of them use proper confidence intervals. The Grok-4 situation earlier this year was a perfect case study: sky-high benchmark numbers, real-world performance that left users deeply unimpressed. Reddit threads with hundreds of comments asking 'why does this feel so much worse than the score suggests?' The benchmark gaming problem is real. Models can be fine-tuned on benchmark-adjacent data. Evaluation setups can be cherry-picked. This is exactly why OSWorld matters more than most: it's hard to quietly overfit on 369 diverse real computer tasks without it being obvious. And it's why production performance on actual desktops matters more than any leaderboard number. Coasty's 82% isn't a press release stat. It's a score on the hardest standardized test for computer use agents that exists, and it holds up in real workflows.

Why RPA Is a Dead End and Most Enterprises Haven't Figured That Out Yet

UiPath's Q1 FY2026 results came out and the company is pivoting hard toward 'agentic automation.' That tells you everything. The old model of RPA, where you hire consultants to painstakingly map every click in a business process and then pray the UI never changes, is dying. It's expensive to set up, brittle when software updates, and completely useless for anything that requires judgment or handles variation. The real cost isn't the software license. It's the maintenance. Every time a vendor updates their web portal, every time an internal tool gets a redesign, someone has to go fix the bot. That's not automation. That's just moving the manual work from the task itself to the task of maintaining the task. A true computer use agent doesn't need a script. It reads the screen like a person does, figures out what to do, and does it. That's a fundamentally different architecture, and the benchmark gap between rule-based RPA and a real computer-using AI agent is not 10 or 20 percentage points. It's the difference between a calculator and a coworker.

Why Coasty Exists

I'm going to be straight with you. I write for Coasty, so take that for what it is. But the reason I'm here is because I watched teams spend real money on Anthropic's computer use API, hit walls with reliability on anything more complex than a three-step workflow, and then go looking for something better. Coasty was built specifically to top OSWorld, the hardest real-world computer use benchmark out there, and it did. 82%. That's not a coincidence. It controls actual desktops and browsers and terminals, not just API endpoints wrapped in a chat interface. The desktop app means you can run it on your own machine. The cloud VMs mean you can scale without touching your infrastructure. The agent swarms mean you can run tasks in parallel instead of waiting for one agent to finish before the next starts. The free tier means you can try it before you spend a dollar. And BYOK support means you're not locked into someone else's model pricing forever. The benchmark score matters because it correlates directly with how often you're going to have to intervene and fix something. At 82%, you're intervening a lot less than at 72% or 75%. That's the whole product.