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
Claude Set a Real Alarm on Android via Intent System — No Hacking, But Transparency Issues Remain
Claude AI used Android's normal intent system to create a native alarm in the Samsung Clock app. The user initially feared a breach, but it's standard app-to-app communication. The lack of upfront disclosure about device-level actions raises transparency concerns.
OpenAI Training Costs Projected to Exceed Anthropic's by 4-5 Times Annually
According to confidential financials reported by the Wall Street Journal, OpenAI expects to spend 4-5 times more on training than Anthropic each year for the next five years. The expense scale is described as mind-boggling.
Claude Code v2.1.128: OTEL isolation, MCP fixes, plugin .zip support, and 20+ bug fixes
Claude Code v2.1.128 stops subprocesses from inheriting OTEL_* env vars, adds .zip plugin support, fixes MCP reconnection flooding, and fixes parallel shell tool cancellation.
Claude API Cost Visibility Concerns for Indie Developers
A Reddit discussion highlights that Claude Sonnet API's lack of granular cost tracking may lead indie developers to drop it despite its quality, with bills of $400–$900 catching them off guard due to insufficient observability compared to AWS-style monitoring.