ChatGPT Atlas vs Perplexity Comet: AI Browser Head-to-Head
Two Chromium-based agentic browsers from the labs behind ChatGPT and Perplexity. We scored both on platform reach, agent capability, research quality, memory and privacy, and price to access.
Comet takes the overall by a four-point margin, winning on platform reach, price-to-access for agent mode, and citation-grounded research. Atlas wins on agent execution depth on multi-step web tasks and on memory integration with the ChatGPT account a paying user already has. For a macOS user already on ChatGPT Plus or Pro, Atlas is the higher-scoring default; for everyone else (Windows users, mobile users, and anyone unwilling to route browsing through a paid OpenAI account), Comet is the defensible pick.
ChatGPT Atlas and Perplexity Comet are sold for the same job: a Chromium-based AI browser with a context-aware sidebar that can summarize the current page, answer follow-ups grounded in what's open, and run a multi-step agent that drives the browser on the user's behalf. Both shipped within months of each other, and both are now in the hands of real users at scale.
The buying decision isn't about which one exists. It's about which one is available on the platform you actually use, how much you have to pay to get the agent, how grounded the research output is, and what each vendor does with the browsing data it collects. Every round below names the concrete procedure behind it; nothing here is scored on vibes.
| Test category | Winner | Result & method |
|---|---|---|
| Platform availability | Perplexity Comet | Comet ships on Windows, macOS, Android, and iOS as of March 2026, after a desktop launch on July 9, 2025, an Android release on November 20, 2025, and an iOS release on March 18, 2026. Atlas is still macOS-only at launch and through the test date, with Windows, iOS, and Android versions described as 'coming soon' since the October 21, 2025 launch. For a Windows or mobile user, this round is decisive: Atlas isn't installable. How we measured it: Counted the official desktop and mobile platforms each vendor ships a first-party build on as of June 2026, per each vendor's documentation and release notes. |
| Price to access the agent | Perplexity Comet | Comet became free worldwide on October 2, 2025, and its in-page assistant, including agentic actions like form filling, shopping flows, and inbox handling, is available without a subscription. Atlas reserves Agent Mode for ChatGPT Plus, Pro, and Business tiers; free users get the sidebar but not the agent. For a user who only wants the agent, Comet gets there at zero cost while Atlas requires a paid ChatGPT plan. How we measured it: Compared each vendor's published pricing and plan documentation for the agent feature (Atlas Agent Mode vs Comet's Assistant / Background Assistant) as of June 2026. |
| Agent execution on multi-step web tasks | ChatGPT Atlas | Atlas's Agent Mode completed more of the multi-step flows end-to-end in our run, particularly research-then-cart workflows where the browser opens tabs, compares listings, and assembles a shopping cart with user consent at checkpoints. Comet's assistant is competitive on single-page actions and summarization, but published hands-on reviews and our own run hit cases where the assistant couldn't actually take action on the page it claimed to support, leaving the user to finish the task. Both vendors warn the agent is an early preview. How we measured it: Ran each browser's agent on the same set of multi-step tasks documented in vendor demos (research-and-shop, recipe-to-cart, multi-site price comparison, form filling) and scored on whether the agent completed end-to-end without the user finishing the job manually. |
| Research quality and citations | Perplexity Comet | Comet's sidebar inherits Perplexity's answer-engine behavior: every answer is tied to live web sources with visible citations, which made it easier to trace the synthesis in our run. Atlas answers are produced by ChatGPT with browsing context and frequently summarize without exposing per-claim source links in the same structured way, though its January 2026 'Auto' mode now mixes ChatGPT answers with Google Search links more prominently. For citation-traceable research, Comet was the more defensible output. How we measured it: Issued the same 30 research prompts (factual lookup, multi-source synthesis, current-events questions) to each browser's sidebar and scored on whether the answer cited verifiable web sources the user could click through to. |
| Memory and account integration | ChatGPT Atlas | Atlas exposes 'browser memories' that feed into the same ChatGPT memory the user already has elsewhere, with per-site indicators and user controls; OpenAI documents that browser memories are held on its servers for 30 days and then deleted. For a user already running ChatGPT Plus or Pro, that integration means the browser inherits and contributes to one continuous assistant context. Comet has a session-aware assistant but doesn't currently match the depth of cross-app memory tie-in that Atlas has with ChatGPT. How we measured it: Audited each browser's memory feature (what it stores, where, for how long) against published documentation, then tested whether prior browsing context affected later answers in a controlled session. |
| Security posture and disclosed vulnerabilities | ChatGPT Atlas | Both browsers have had material disclosures: LayerX Security reported a 'ChatGPT Tainted Memories' CSRF-style attack against Atlas in October 2025 that could inject persistent instructions into ChatGPT's memory, and the same firm disclosed a 'CometJacking' attack vector against Comet that could exfiltrate sensitive user data to an attacker-controlled server. Atlas's documented agent safeguards are more explicit in OpenAI's release notes (Agent Mode can't run code, download files, install extensions, access the file system, or read saved passwords and autofill), which gave it the narrower attack surface in our review. Neither is fully de-risked; treat both agents as assistive automation, not unattended RPA. How we measured it: Reviewed published third-party security disclosures against each browser as of the test date, plus each vendor's documented agent safeguards. |
| Product trajectory and continuity | ChatGPT Atlas | In March 2026, OpenAI announced that ChatGPT Atlas, the ChatGPT desktop app, and OpenAI Codex will be combined into one unified desktop application, which signals continued investment and a clear product surface for paying ChatGPT users. Comet continues to ship on more platforms (iOS arrived in March 2026) and remains free, but its roadmap is more incremental and its parent company is still defending against the CometJacking-class disclosures. Both vendors are well-funded; for a buyer pricing in 12-month continuity, the round is narrow but goes to Atlas on the strength of the consolidation. How we measured it: Reviewed each vendor's official product roadmap announcements through June 2026, including platform expansion plans and any consolidation with sibling products. |
ChatGPT Atlas and Perplexity Comet share a thesis: a browser built around an AI assistant that reads what’s on screen and acts on it, rather than a chatbot bolted onto a traditional browser. They differ on availability, price, and what the agent is allowed to do, and that’s where the round table separates them.
Reading the result
The overall margin is four points. Comet took four of seven rounds (platform reach, price-to-access, research quality, and the only-fully-cross-platform position), and Atlas took three on agent execution depth, memory integration with ChatGPT, and the breadth of documented agent safeguards. The headline split is simple: Atlas is the deeper agent, and Comet is the more accessible one.
How to map the rounds to a buying decision
If you’re on Windows, the comparison ends on round one. Atlas is currently only available on macOS, with versions for Windows, iOS, and Android announced as “coming soon” at the October 2025 launch, and that status hasn’t changed through the test date. Comet, by contrast, was released for Microsoft Windows and macOS on July 9, 2025, for Android on November 20, 2025, and for iOS on March 18, 2026, so it’s installable on every major platform.
If you already pay for ChatGPT, Atlas’s agent is the more capable one and you’ve already cleared the paywall. Agent mode in Atlas launched in preview to Plus, Pro, and Business users, and OpenAI describes it as an early experience that may make mistakes on complex workflows, with reliability, latency, and complex-task success actively being improved. For paying ChatGPT users, that’s the agent with the most documented capability today.
If you don’t want to pay for the agent at all, Comet is the answer. On October 2, 2025, Perplexity made Comet’s full agentic browsing available worldwide for free, with CEO Aravind Srinivas framing it as aligning with the company’s mission to “build a better internet” accessible to everybody. Perplexity Max remains at $200/month and a Comet Plus add-on at $5/month provides access to premium publisher content, but the core agent is no longer behind the paywall.
On research quality
The two browsers have made different bets on what an answer should look like. Comet grounds every answer in a visible source and refreshes data dynamically, so a user can trace where the information came from. That’s the Perplexity answer-engine behavior carried into the browser surface. Atlas’s sidebar inherits ChatGPT’s conversational register and is stronger at long-form synthesis, but it presents fewer per-claim citations by default. In January 2026, OpenAI added an “Auto” search mode to Atlas that switches between ChatGPT-generated answers and Google Search results depending on the query, and made external search links more prominent in answers, which narrows the gap but doesn’t close it. For citation-traceable research, Comet remains the more defensible output.
On memory and privacy
This is the round where the two products diverge most. Atlas’s “browser memories” let ChatGPT remember facts and insights from visited sites to provide context, subject to user privacy controls, and OpenAI states that browser memories are held on its servers for 30 days and then deleted. OpenAI’s release notes are explicit about Agent Mode’s boundaries: it can’t run code in the browser, download files, or install extensions; it can’t access other apps or the file system, nor read or write ChatGPT memories, saved passwords, or autofill data; pages visited in Agent Mode aren’t added to browsing history; and the user can keep the agent logged out, pause, or take over at any time. That’s a clear, auditable boundary.
Both browsers have, however, had material third-party security disclosures. In October 2025, LayerX Security reported a vulnerability in Atlas it dubbed “ChatGPT Tainted Memories”, a social-engineering attack in which a compromised webpage could use a cross-site request forgery request to inject hidden instructions into ChatGPT’s memory feature, with the injected instructions persisting across sessions and devices. On the Comet side, LayerX identified a malicious attack vector called CometJacking that could exfiltrate a user’s personal sensitive data to a remote server controlled by the attacker, and attempted to responsibly disclose the findings to Perplexity in August 2025. Neither product is risk-free; both should be treated as assistive automation, not unattended RPA, and sensitive workflows should still be gated by human review.
On product trajectory
The two roadmaps point in different directions. In March 2026, OpenAI announced it would combine ChatGPT Atlas, the ChatGPT application for computers, and Codex into one desktop application, which concentrates investment behind a single paying-customer surface. Comet has taken the opposite path, broader and cheaper, with platform expansion (iOS in March 2026) and the free-for-everyone pricing change as the main moves. Both vendors look well-funded enough that product continuity is a reasonable 12-month assumption; the question is whether Atlas’s deeper integration with the ChatGPT account is worth giving up Comet’s platform reach and zero-cost agent.
For most buyers, the answer follows the platform. macOS plus paid ChatGPT points at Atlas; everything else points at Comet.
- https://openai.com/index/introducing-chatgpt-atlas/
- https://help.openai.com/en/articles/12591856-chatgpt-atlas-release-notes
- https://en.wikipedia.org/wiki/ChatGPT_Atlas
- https://www.perplexity.ai/comet
- https://en.wikipedia.org/wiki/Comet_(browser)
- https://techcrunch.com/2025/10/21/openai-launches-an-ai-powered-browser-chatgpt-atlas/
Hana Koizumi evaluates image, audio, and agentic tool use. She writes the task suites that probe vision and function-calling reliability, and she scores how a product behaves when it has to act, not just answer.