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

The Stanford Numbers Everyone Is Misreading

The 2026 Stanford HAI AI Index report dropped in April and it confirmed what people in the computer use space already knew: OSWorld is now the most important benchmark in AI, and the progress over the past 12 months has been genuinely wild. Agents went from roughly 12% task success to around 66% for leading models in one year. That's a 5x jump. On SWE-bench Verified, coding performance went from 60% to near 100% in a single year. These are not small moves. But Stanford's own researchers flagged something that got buried in the hype cycle: benchmarks may not map to real-world results. That's a polite academic way of saying some of these scores are inflated. OSWorld tests 369 real computer tasks across operating systems including spreadsheet work, browser navigation, file management, and terminal commands. It's the closest thing we have to a fair fight. But when labs optimize specifically for a benchmark, scores go up without real-world capability following. IEEE Spectrum quoted Stanford researchers warning about exactly this. So when you see a score, the first question isn't 'how high is it.' It's 'how did they get there.'

Where Every Major Player Actually Stands

• ●OpenAI's Computer-Using Agent (CUA) launched in January 2025 with a 38.1% OSWorld score. That was their headline number. They called it 'state of the art.' It was, briefly, for about six weeks.
• ●Claude Sonnet 4.5 hit 61.4% on OSWorld by September 2025. Real progress. Anthropic deserves credit. But their own blog noted it's still 11 percentage points below what a regular human scores on the same tasks.
• ●Simular Agent S2 posted strong OSWorld results in early 2025 using a modular generalist-specialist architecture. Interesting approach. Promising research. Not yet a product most teams can run in production.
• ●Stanford's AI Index confirmed the overall field average crept to roughly 66% for leading agents by late 2025. The average. Meaning half the agents people are actually using score below that.
• ●Coasty sits at 82% on OSWorld. That's not a rounding error. That's a double-digit gap over Claude Sonnet 4.5 and a 44-point gap over where OpenAI started the year. It's the highest publicly verified score on the leaderboard.
• ●Human baseline on OSWorld is approximately 72-74%. Coasty at 82% means it's already outperforming the human benchmark on the tasks OSWorld tests. That's the number that should be making headlines.

The Benchmark Gaming Problem Nobody Wants to Talk About

Here's the uncomfortable truth about AI benchmarks in 2026: almost every frontier lab reports capability benchmark results, according to Stanford's own index. Almost every one of them also has an incentive to make those numbers look as good as possible. The OECD flagged benchmark gaming as a 'particularly significant challenge' in their 2026 AI trajectories report. When you train a model with OSWorld tasks in the training data, or when you fine-tune specifically to ace the test set, your score goes up without your agent actually getting better at the job. It's teaching to the test, and it's rampant. This is why deployment horror stories keep piling up. Companies buy into a benchmark score, stand up the agent in production, and watch it fall apart on anything slightly outside the test distribution. RPA vendors like UiPath built entire empires on demo-ware that looked great in controlled environments and broke constantly in the wild. The AI agent wave risks repeating that exact mistake. The only real defense is transparent, third-party verified benchmarks on diverse task sets, which is exactly why OSWorld matters more than any lab's internal eval.

Why Your Company Is Still Paying Humans to Do Robot Work

Let's talk about what's actually at stake here, because the benchmark debate is fun but the business case is what should be keeping executives up at night. Parseur's 2025 research found that manual data entry costs U.S. companies $28,500 per employee per year. Clockify's research puts the share of work time spent on repetitive tasks at 62%. Smartsheet found over 40% of workers spend at least a quarter of their week on manual work. And 56% of those workers report burnout from it. These aren't edge cases. This is the default state of most knowledge-work organizations right now. The tools to fix this exist. Computer use agents that can navigate real desktops, fill forms, pull data from legacy systems, run terminal commands, and execute multi-step workflows across applications are here. They work. The gap between 'benchmark score' and 'actually deployed and saving money' is where most companies are stuck, and it's mostly a vendor selection problem, not a technology problem.

Why Coasty Exists and Why 82% on OSWorld Is the Only Number That Matters Right Now

I've looked at a lot of computer use agents. I've run tasks through Claude's computer use API. I've tested Operator. I've read every OSWorld leaderboard update for the past year. And the honest answer is that most of them are impressive demos that become frustrating production tools the moment your workflow gets even slightly complicated. Coasty was built differently, and the 82% OSWorld score is the proof, not the pitch. That score is independently verified on the full OSWorld task set, not a cherry-picked subset. It covers real operating system tasks across browsers, terminals, spreadsheets, and file systems. It beats the human baseline. Nothing else on the public leaderboard is close. But the score isn't why I'd recommend it. I'd recommend it because it controls real desktops and browsers, not just API endpoints. It runs cloud VMs so you don't need to provision your own infrastructure. It supports agent swarms for parallel execution, which means you can run 10 workflows simultaneously instead of waiting for one to finish. It has a free tier so you can test it without a procurement process. And it supports BYOK so you're not locked into one model provider. That combination of benchmark performance and actual production architecture is rare. Most tools have one or the other. Coasty has both, and the gap between it and the next competitor on OSWorld is not closing fast.