Google is restructuring the team behind Project Mariner as industry attention shifts away from standalone browser agents toward coding and terminal-based agents. The signal is partly numerical: Perplexity’s Comet reached only 2.8 million weekly active users in December 2025, while OpenAI’s ChatGPT Agent reportedly fell below 1 million weekly active users. Against ChatGPT’s weekly audience in the hundreds of millions, browser-agent usage is statistically marginal, indicating a major gap between industry hype and consumer adoption.

The performance trend also favors terminal-first agents. Browser agents rely on repeated screenshots and visual interpretation, which raises compute cost, slows execution, and reduces reliability. By contrast, terminal-based systems such as Claude Code and OpenClaw operate in text, and one industry estimate suggests they require roughly 10x to 100x fewer steps to reach the same result. This efficiency shift helps explain why Google is folding Mariner capabilities into broader agent products rather than treating browser navigation as the primary product category.

The market is not moving to a pure terminal model, however. One startup claims a video-based computer-use model is 50x more efficient than earlier screenshot-based approaches, suggesting GUI automation is still improving. Another estimate in the article proposes an 80/20 split: many tasks can be solved through the terminal, but a persistent minority still requires GUI access, especially for legacy software and websites without APIs. The broader statistical pattern is therefore strategic reallocation, not full abandonment: AI labs are prioritizing coding agents while preserving computer-use capabilities as a secondary but necessary layer.