Coasty AI Agent Solves CAPTCHA Challenges Up to Level 6 Without Training
Coasty's Computer Using Agent Handles Real Desktop Challenges
Coasty's Computer Using Agent (CUA) has demonstrated the ability to solve CAPTCHA challenges up to Level 6 without being specifically trained for "I'm not a robot" tests. The agent achieved 82% on the OSWorld benchmark, which represents state-of-the-art performance for computer-use agents operating in real desktop environments.
The agent handles various web interface challenges that typically break other agents, including:
- CAPTCHA challenges up to Level 6
- Browser popups
- Cookie banners
According to the source, the developers did not teach the CUA to solve "I'm not a robot" challenges specifically, noting that "the irony is not lost on us." The agent's performance suggests it has developed generalized computer interaction capabilities rather than specialized solutions for individual challenge types.
A replay link is available for those interested in seeing the agent in action: https://coasty.ai/share/1cd404ae-3fcb-4d7f-b9d4-dac7aa26fc6d
📖 Read the full source: HN AI Agents
👀 See Also
Opus 4.6 Medium vs Low: Performance Differences and Pricing
Opus 4.6 medium costs approximately 50% more than the low version but addresses significant laziness issues found in the low-powered model. The medium version sits between low and high in performance benchmarks.
OpenClaw v2026.7.1: Control UI Overhaul, Onboarding, Mobile Apps, GPT-5.6, Tencent Hy3, Meta Muse Spark 1.1
OpenClaw v2026.7.1 brings a major Control UI overhaul, redesigned onboarding, updated iOS/Android/macOS apps, GPT-5.6 compatibility, Tencent Hy3 and Meta Muse Spark 1.1 support, and improved Codex and coding-agent workflows.
MiMo-V2.5-Pro Benchmarked: Strong Social Deduction Reasoning, Good Value vs K2.6
MiMo-V2.5-Pro competes with Kimi K2.6 in autonomous Blood on the Clocktower games, with a lopsided 88% Good / 48% Evil win rate, costs $0.99/game at 183k output tokens, and is practical with 2-3 hour matches.
Claude Max $100 subscription usage data for API extension task
A Claude Max $100 subscription user reports consuming 13% of a 5-hour session to extend an existing API with favorite library functionality, with context usage at 11% and weekly usage increasing from 5% to 6%.