Skip to the content.

2026 (rolling) — AI browsers: ChatGPT Atlas, Perplexity Comet, and the agentic-browser category

Why this matters

In late 2025, two AI-first browsers shipped within months of each other: Perplexity Comet (July 2025) and ChatGPT Atlas (October 21, 2025). By 2026 the category had a name — agentic browsers — and a defining feature: the browser is no longer a tool the agent uses, the agent is a tool the browser is built around.

This entry exists because the agentic-browser category is structurally adjacent to Claude Code in a specific way: it’s the same “agent that drives a controlled environment” pattern, applied to the web instead of to a codebase. The parallels with Webwright (Microsoft Research’s CC-as-browser-agent skill, see the plugin context across the skills entry) are explicit. The agent-controlling-a-browser shape is one of the highest-leverage shapes in 2026, period.

The two main agents

ChatGPT Atlas

Launched 2025-10-21 by OpenAI. Chromium-based browser with ChatGPT integrated at its core. Initially macOS-only; in March 2026, OpenAI announced Atlas would be merged with ChatGPT and Codex into a single desktop superapp — Windows release timeline unclear.

Architecturally a Chrome fork (familiar UI, tab model, extensions) with ChatGPT as the default search engine and as an omnipresent agent available on every tab. Agent functionality requires ChatGPT Plus ($20/mo) or higher.

The headline use-case Atlas pitches: autonomous web research. Open a tab, give Atlas a task (“compare these five SaaS vendors on pricing, integrations, customer reviews”), watch it navigate through tabs, take notes, surface conclusions. The same shape as a human research workflow, autonomously.

Perplexity Comet

Launched 2025-07 by Perplexity. Same agentic-browser category. Comet went from $200/month exclusivity to completely free in October 2025 — when Perplexity removed the waitlist and made it accessible to all users at no cost. The pricing move is itself the story: when an AI product flips from $200/mo to free, the platform play is the strategic move, not the per-user revenue.

Comet’s specific pitch: proactive AI assistants that guide users toward outcomes rather than links. Designed as AI-first; replaces traditional search-driven interactions entirely.

What “agentic browser” means as a category

Three properties define the category, per the 2026 comparison writeups:

  1. The agent is always available. Not a “click this button to summon AI” surface; the AI is a persistent sidebar / chat / floating action that follows the user across tabs.
  2. The agent can act on the browser. Click links, fill forms, navigate, scroll, copy content. This is the structurally different bit from “ChatGPT in a browser tab” — the agent isn’t inside the page, it’s driving the browser.
  3. Cross-tab context. What you read on one tab informs what the agent does on the next. The browser becomes a state surface the agent uses, similar to how a filesystem is a state surface for Claude Code.

The adjacent technology stack: Claude Computer Use (the Anthropic API capability that lets a model click and type through a desktop interface) is the underlying primitive that makes this category technically feasible. Atlas, Comet, and similar tools wrap this primitive into a product surface.

The Webwright comparison

Microsoft Research’s Webwright is the DIY version of the same idea — a Playwright workspace the agent (Claude Code, Codex, OpenCode, Hermes, Pi) drives one bash command at a time. The Pinggy guide to AI harnesses framing applies here: Atlas / Comet are productised agentic browser harnesses; Webwright is the DIY harness for the same work.

For someone who’d otherwise spend $20/mo on Atlas or use Comet free: Webwright + Claude Code gives you the same agentic- browser capability inside your own agent harness, with full control. The tradeoff: Webwright runs Playwright (a real browser engine) under your harness; Atlas / Comet run their own UX-tuned product. Different shape, same primitive.

What this means for Claude Code use

Three durable claims:

  1. Browser-driving agency is a 2026-mainstream pattern. Up until late 2025 this was research / demo territory; from Q4 2025 onward it’s a product category with multiple players. The technical primitive (Claude Computer Use, OpenAI’s equivalent) is mature enough that productisation can compete on UX, not on whether-it-works.
  2. The browser is becoming an “MCP server” in shape. Per the MCP entry: MCP servers expose tools the agent can call. An agentic-browser product is the browser-as-a-tool-surface — the agent’s primary verb is “navigate to a URL and do something.” This is a category- level convergence: agents, regardless of their wrapping, end up driving similar primitives (filesystem, browser, shell).
  3. The merger announcement matters. OpenAI’s March 2026 announcement of merging Atlas + ChatGPT + Codex into a single desktop superapp is the commercial version of what Anthropic shipped with managed agents — different surfaces (browser, chat, CLI) become different views of the same agent infrastructure. Expect Anthropic-side equivalents to follow.

How the AI browsers compare

Browser Released Free? Best for
Perplexity Comet July 2025 Yes (since Oct 2025) Perplexity users; research-heavy work
ChatGPT Atlas Oct 2025 $20/mo for agent features ChatGPT users; mac-only as of Feb 2026
Arc Browser (legacy) Pre-2025 Yes Now in maintenance mode; the ‘AI features’ were retrofit
Dia (The Browser Company) 2025 Beta access Arc’s successor; AI-first design

The 2026 community consensus: Atlas and Comet are great for their respective ecosystems; pick based on whether you use ChatGPT or Perplexity as your primary chatbot. The category itself is moving faster than any browser cycle since Chrome launched — expect feature parity between leaders to shift quarterly.

Try this

If you spend any meaningful time on browser-based research:

  1. Try Comet for a week. It’s free; no commitment beyond downloading. Pick a research task you’d normally do across 30+ tabs. Watch what Comet does differently.
  2. Compare against Webwright + Claude Code doing the same task. The browser-agent shape is the same; the harness wrapping is different. Which one fits your workflow better says something about whether you want browser-as-product or browser-as-tool.
  3. Read the AI browser landscape overview for the broader context — Arc, Dia, and the pre-AI-browser wave that’s now retrofitting.

The point isn’t to switch browsers. It’s to recognise that “agent drives a controlled environment” is a 2026 shape that productises across many surfaces — codebases (Claude Code), web (Atlas, Comet), 3D modelling (Claude Design’s Autodesk connector), audio (Ableton connector), finance workflows (Anthropic finance agents). The harness changes; the underlying pattern is the same.

Sources