I am Claude-pilled again, but not for the reason most people think. It is not that the harnesses have finally caught up to model capabilities; it is the exact opposite. Anthropic built Claude Code as the most powerful harness possible, and now we are seeing the models grow into the capability of that harness. Claude Opus 4.5 feels almost perfect at navigating the environment it was designed for. In my testing, Claude Code plus Opus 4.5 is performing better than Antigravity and Opus 4.5, especially on tasks that require fewer iterations and higher code quality.
The Performance Gap: Claude Code vs Antigravity
While both systems are capable, the difference lies in efficiency. When I ask for a new feature, Claude Code delivers accurate results while Antigravity takes longer due to less efficient tools. The persistence factor is also higher. When I was building Button Bench to see if LLMs would press a restricted red button, Claude Code kept iterating on adversarial prompts until it achieved a working result. It simply refuses to quit until the job is done. This level of reliability is also why I have moved away from Codex, which feels unusably slow in comparison.
The Browser Agent and Task Automation
The Claude Browser Agent is the ultimate companion to this workflow. It takes tasks that have been on my to-do list for weeks and completes them with minimal prompting. By executing multi-step workflows like form filling and research synthesis directly in Chrome, it handles the administrative friction that engineers usually avoid. As I noted in Best LLMs 2025 Comparison, the wrapper matters immensely. In this case, the wrapper is an agent that can actually use the web like a human does.
Capabilities Over Interface
The Claude web interface is still not particularly special compared to ChatGPT features. However, the actual work these agent harnesses can do is incredible. The model is essentially a match for the terminal-first CLI agent, enabling deep codebase reasoning and self-correction. We are seeing a shift where raw model power is meeting high-intent tool use. For more on this trajectory, you can see my thoughts in 2025 AI Timeline. This ecosystem is now about capability over polish, and it is winning.
