AI Browser Agents in 2026: What Comet, Atlas, and Copilot Studio Actually Do Differently

“AI browser agent” has become a category with five or six genuinely different products under it, and they don’t do the same job. Some are entire browsers built around an assistant. Some are extensions bolted onto Chrome. One is an enterprise workflow platform feature with governance controls the others don’t have. If you’re deciding whether to use one, or which one, the differences matter more than the shared marketing language suggests.

The two shapes this category takes

The clearest split isn’t by vendor, it’s by architecture. Perplexity Comet and ChatGPT Atlas are standalone, AI-native browsers: the agent is the primary interface, and you browse through it rather than adding it to an existing tool. Comet has been free since March 2026, with a paid Comet Max tier for heavier use. Atlas launched with a freemium structure: free tier, Plus at $20/month, Pro at $200/month, and a Business plan at $24 per user per month, with Agent Mode unlocked starting at the Plus tier.

The other shape bolts automation onto a tool you already use. Claude in Chrome is a browser extension, available on Anthropic’s paid Claude plans starting at Pro ($20/month), that gives Claude direct control over your existing Chrome session rather than asking you to switch browsers. As of July 2026, the related Claude Cowork product also runs sessions remotely on Anthropic’s servers, extending the same agent capability to web and mobile. Microsoft’s Copilot Studio Computer Use, generally available since May 13, 2026, takes this further into enterprise territory: it’s a feature inside a workflow automation platform, priced through Copilot Studio’s credit-based system, built specifically for regulated environments.

What happened to Google’s entry

If you were tracking Google’s Project Mariner, the short version is that it doesn’t exist as a standalone product anymore. Google shut it down on May 4, 2026, after about 17 months of development, and folded the underlying technology into two places: Gemini Agent, the task-automation layer inside Google’s main assistant, and Chrome’s Auto Browse feature, which handles multi-step web navigation directly inside the browser you’re probably already using. That’s worth knowing before you go looking for a Mariner sign-up page that no longer exists.

What the benchmarks actually say

Two numbers are worth anchoring to, and it’s worth being honest that they’re not directly comparable to each other since they measure different things.

On WebVoyager, a 643-task benchmark of real-world browsing tasks, the open-source Browser Use project running on GPT-4o hit 89.1% success, edging out OpenAI’s own Operator at 87%. That’s a genuinely notable result: an open-source tool built on a general-purpose model outperforming a purpose-built commercial agent on the same benchmark, which says more about how much of browser-agent performance comes from orchestration and tool design than from the underlying model alone.

Separately, Claude Sonnet 4.6 reached 72.5% on OSWorld in February 2026, a benchmark that tests broader OS-level task completion rather than pure web browsing, and that score has been described as roughly matching average human performance on that specific benchmark. It’s a different task category, so don’t read it as “worse than Browser Use,” it’s answering a different question about what the agent can do beyond the browser tab itself.

Choosing one, practically

Agent	Vendor	Shape	Starting price	Where it’s strong
ChatGPT Atlas / Agent Mode	OpenAI	Standalone browser	Free tier; Agent Mode from $20/mo	Broad task coverage, tight ChatGPT integration
Perplexity Comet	Perplexity	Standalone browser	Free since March 2026	Search-native workflows
Claude in Chrome / Cowork	Anthropic	Browser extension + remote agent	From $20/mo (Claude Pro)	Existing Chrome workflows, cross-device via Cowork
Gemini Agent / Chrome Auto Browse	Google	Built into Chrome and Gemini	Included with Google AI plans	Native Chrome integration (successor to Project Mariner)
Copilot Studio Computer Use	Microsoft	Enterprise workflow platform	Credit-based, from ~$200/mo	Governance: DLP, audit trails, human-in-the-loop
Browser Use	Open source	Self-hosted framework	Free (self-hosted)	Highest measured WebVoyager score (89.1% on GPT-4o)

For a small team automating a handful of repetitive browser tasks, an extension you already have access to through an existing subscription is the lowest-friction start. For anything touching client data, financial systems, or regulated workflows, weight governance over raw benchmark score, which is why Microsoft’s enterprise-focused offering is worth the evaluation even if it’s not the top scorer on any public leaderboard. And for teams comfortable building their own guardrails, the open-source option currently posts the best measured result on a real benchmark, at zero licensing cost.

None of these tools are at the point where “fully autonomous, no supervision” is a safe default. If you’re integrating one into a client-facing product rather than using it as an internal productivity tool, that’s a meaningful build decision worth scoping carefully rather than bolting on as an afterthought, the same way we’d approach any AI agent orchestration work: start supervised, expand autonomy only where the failure cost is genuinely low.