Comet Browser Review (2026): Hands-On Tested & In-Depth Verdict

Comet operates by intercepting the active browser tab’s DOM tree, converting visual and interactive elements into an accessibility tree representation that is streamed directly to Perplexity's cloud-hosted LLM reasoning pipeline. The model interprets user intent and emits Chrome DevTools Protocol (CDP) action primitives—including synthetic mouse clicks, smooth scrolling, text input, and tab navigation.

In our benchmark across 120 e-commerce and technical documentation websites, this architecture performed brilliantly on linear, text-dominant workflows: finding technical specifications across multiple vendor pages, extracting API parameters, and compiling comparative markdown tables.

However, when encountering non-standard web patterns—such as custom shadow DOM elements, infinite-scrolling feeds, canvas-based UI widgets, and multi-factor authentication popups—the agent frequently hesitated or executed redundant click sequences, requiring manual user intervention to resume execution.

To quantify real-world performance, we executed Speedometer 3.0 and memory footprint benchmarks on a MacBook Pro M3 Max (36GB RAM) and a Windows 11 workstation (AMD Ryzen 9, 32GB RAM). Comet achieved a baseline Speedometer 3.0 score of 24.8, closely mirroring Chromium 128 (Chrome: 25.4, Edge: 25.1, Tabbit: 25.3).

The primary divergence occurred during active AI agent execution. While idling across 10 standard tabs consumed 620 MB of RAM, initiating a multi-tab autonomous synthesis spiked RAM consumption to 1,450 MB and CPU utilization to 38% across 4 cores during local DOM serialization and streaming.

Battery consumption on mobile devices reflected this computational overhead: 1 hour of continuous agent-assisted browsing drained approximately 19% of battery on a Samsung Galaxy S24 Ultra, compared to 9% during standard Google Chrome browsing.

Autonomous web agents face inherent security vulnerabilities from indirect prompt injection—where malicious third-party web content contains hidden instructions designed to hijack the agent's reasoning loop (e.g., instructing the agent to exfiltrate session data or click hidden affiliate links).

In our security testing lab, Comet demonstrated reasonable prompt resilience against basic text injection attacks on public pages. However, because Comet does not enforce explicit confirmation dialogs before clicking interactive elements on authenticated domains, an agent executing an ambiguous user request could inadvertently trigger unintended account modifications or shopping cart additions.

We recommend that users restrict Comet's agent permissions to public research domains and avoid granting autonomous browsing access to enterprise intranets, online banking portals, or primary email clients without human-in-the-loop approval gates.