Perplexity Comet in 2026: The AI Browser That Browses for You — an Honest Review
You ask Perplexity Comet to order a gluten-free cheese pizza from your local spot on Seamless. It finds the restaurant, navigates the menu, adds a regular cheese pizza to your cart, types “please make it gluten-free” in the special instructions box, and adds extra cheese you never asked for. Then it tells you it cannot complete the purchase for security reasons.
This is Comet in mid-2026: genuinely impressive in concept, frequently useful, and still tripping over its own feet on anything that requires actual judgment.
What Perplexity Comet actually is
Comet is not a chatbot bolted onto a browser. It is a Chromium-based desktop browser for macOS and Windows, with a Perplexity-powered AI agent built into its core architecture. The assistant lives in a sidebar, sees your open tabs, reads web pages, and can click, scroll, type, and navigate on your behalf.
Perplexity launched Comet in July 2025 as an invite-only product gated behind its Pro subscription. By October 2025, the wall came down. Comet is now free for everyone with a Perplexity account. A $5/month “Comet Plus” add-on unlocks premium news content from publishers like CNN, Fortune, and The Washington Post. It is Perplexity’s answer to Apple News Plus.
The timing is not an accident. Google lost a major antitrust case in 2025 and may be forced to sell Chrome, which commands roughly 72% of the global browser market. Apple received $20 billion per year to keep Google Search as Safari’s default, an arrangement now under active renegotiation. Samsung is reportedly exploring pre-loading Perplexity on future Galaxy phones. When the foundations of the browser market are shifting, releasing a credible alternative is a rational move, not a vanity project.
The AI browser is no longer a novelty
Comet is not alone. The past 18 months have produced a wave of AI-native browsers:
| Browser | Maker | AI Approach |
|---|---|---|
| Comet | Perplexity | Integrated AI agent with sidebar, agentic browsing |
| ChatGPT Atlas | OpenAI | Full ChatGPT integration with agent mode |
| Dia | Browser Company (Arc) | Lightweight AI-centered browser, minimalist |
| Edge | Microsoft | Copilot side panel, enterprise-focused |
| Opera AI | Opera | Rebranded Aria assistant, embedded chat |
Comet’s thesis is that AI should not be a feature you summon. It should be the browser’s operating system. The assistant watches what is open, tracks context across tabs, and can act without you switching tools.
Key features: what works and what does not
Page summarization: reliable
Click the three-line icon next to the address bar. The assistant generates a short summary of whatever page you are on. This is Comet’s most polished feature. It is fast, responses arrive in under a second, and it cites sources.
But summaries are only as useful as their granularity. Comet tends to treat all information as equally important. When Mashable tested it on a Hard Fork podcast episode where OpenAI’s Sam Altman and Brad Lightcap made a surprise entrance, Comet’s summary did not mention the entrance at all. It described a tense, specific exchange as “humorous commentary.” If you need the highlights, Comet works. If you need the interesting highlights, you still need to read.
Agent mode: half brilliant, half broken
Toggle Agent Mode and Comet can act semi-autonomously: find products, compare prices, fill forms, book meetings. It shows what it is doing in real time. A blue frame highlights elements it is interacting with, and a “Steps” tab reveals its chain-of-thought reasoning.
Simple tasks work. Ask it to “find recent reviews of electric bikes, shortlist the top three under $2,000, and open them in tabs.” It executes cleanly.
Complex tasks fail at the edges. Mashable’s Cecily Mauran tested gluten-free pizza ordering. Comet could not find the gluten-free option because it was buried in a toppings dropdown rather than listed as a separate crust type. Instead of reporting the limitation, it added a regular pizza and typed “please make it gluten-free” in the special instructions box. It also added extra cheese unprompted.
Cybernews tested a cat food comparison. Comet found high-meat options and added items to a cart, but did not check every available product. When asked to find the cheapest shop, it analyzed only Google result snippets, not full pages. It missed the deeper answer entirely.
The pattern is consistent across reviews: Agent Mode succeeds on linear, well-structured tasks and fails when the correct action requires understanding an interface’s idiosyncrasies.
AI sidebar and cross-tab context: useful but inconsistent
The sidebar assistant can reference open tabs using @tab1, @tab2 syntax. This lets you compare products across tabs or ask questions that span multiple pages without copy-pasting.
In practice, tab tagging is temperamental. Cybernews found it “inconsistent or a bit buggy.” Second Talent’s review noted that the assistant “watches what’s open, keeps track of your tab context, and can join your workflow midstream,” but the experience depends heavily on what you are asking it to do.
Voice assistant: functional but secondary
Voice input is powered by Whisper-like speech models. You can speak commands or ask questions about the current page. In testing, voice mode performed worse than text-based interaction, a gap PCMag noted as a real limitation.
Auto tab organization: surprisingly good
Ask Comet to group your open tabs by topic and it does it. This is not a headline feature, but it is one of the few that consistently works without caveats. It can go overboard. Mashable noted that it created topic labels for individual tabs that did not fit any category, like a calendar page, adding visual clutter. But ungrouping is a single click.
Memory: the weakest link
Comet advertises “thread memory” that recalls past interactions across sessions. Cybernews tested it directly: after adding cat food to a cart and closing the browser, they asked the assistant to “find the last thing I added to the basket and add it again.” Comet replied that it could not find any recent basket activity.
When the reviewer specified the exact shop, it worked. Contextual memory exists, but it is not reliable enough to depend on.
Privacy: three modes, one ambiguity
Comet offers three privacy tiers:
| Mode | Behavior |
|---|---|
| Standard | Shares contextual data with Perplexity servers to improve results |
| Strict | Limits what the AI can see or remember within a browsing session |
| Local | Runs tasks on-device without external data transmission |
On paper, this is better than Chrome’s binary incognito/normal model. But Comet’s privacy disclosures include the line “anonymized interaction data may still be used for quality assurance.” This means none of the modes are truly zero-data. Without independent audits, it is hard to verify how each mode behaves in practice.
The privacy pitch is directionally correct. Tiered controls give users more agency than Chrome or Edge offer by default. But anyone treating Local mode as air-gapped is making an assumption, not an observation.
Security: the CometJacking wake-up call
Shortly after launch, researchers discovered a class of attacks dubbed “CometJacking.” Malicious web prompts could inject hidden instructions into pages, tricking Comet’s agent into sharing snippets of user data or performing unintended actions.
Perplexity responded with patches and added a visual confirmation layer before the agent executes sensitive tasks. Independent audits from Brave Software and Guardio Labs confirmed that Comet’s sandboxing improved post-launch. But the episode exposed an architectural tension: the more autonomous the agent, the larger the attack surface.
This is not a Comet-specific problem. Every agentic browser, Atlas included, faces the same class of threats. The question is not whether vulnerabilities will be found, but how fast they get patched.
Performance
Comet loads pages as fast as Chrome, which is unsurprising since they share the Chromium engine. The AI summarization engine responds in under a second for most tasks. In side-by-side tests, Second Talent found Comet delivered answers faster than Atlas for basic queries, though Atlas produced more nuanced reasoning on longer conversations.
The main performance complaint is battery drain. Running agent mode and local processing simultaneously consumes noticeably more power than Chrome on laptops. The gap is not drastic enough to be a dealbreaker, but it is real.
Comet vs. the alternatives
| Feature | Perplexity Comet | ChatGPT Atlas | Google Chrome | Microsoft Edge | Arc / Dia |
|---|---|---|---|---|---|
| Base engine | Chromium | Chromium | Chromium | Chromium | WebKit fork |
| AI assistant | Built-in sidebar + voice | Full ChatGPT integration | Gemini testing | Copilot panel | Co-Pilot |
| Agentic tasks | Full agent mode | Multi-step actions | None | Limited | Semi-agentic |
| Context memory | Session + thread | Browser memory | History only | Limited | Local notes |
| Privacy controls | Standard/Strict/Local | Site visibility toggle | Incognito only | InPrivate mode | Local-first |
| Chrome extensions | Full support | Full support | Supported | Supported | Limited |
| Price | Free (+ $5/mo Plus) | Free | Free | Free | Free |
| Best for | Researchers, analysts | Writers, AI power users | General browsing | Enterprise | Designers |
Comet leads on agentic browsing and integrated AI. Atlas leads on reasoning depth. Chrome and Edge lead on stability and ecosystem maturity. Arc and Dia lead on design. There is no clear overall winner. The right choice depends on what you actually do in a browser.
What nobody tells you
The browser is distribution for the search engine, not the other way around. Perplexity is an AI search company. Comet defaults to Perplexity search. Chrome defaults to Google Search. Google built a trillion-dollar business on this relationship. Perplexity is attempting the same playbook with an AI-native stack.
You will spend as much time verifying as you save. Every reviewer across Mashable, Cybernews, Second Talent, and Lifehacker reached the same conclusion: Agent Mode’s output requires double-checking. For simple research tasks, the time arithmetic works out. For anything involving money, scheduling, or important data, the verification overhead erases the automation gain.
The memory feature that does not work is a bigger problem than it looks. Contextual memory is the foundation for everything Comet promises: persistent assistance that learns how you work. If the browser cannot reliably remember what you did yesterday, the “agent that travels the web with you” pitch collapses into “a chatbot sidebar that sometimes references your tabs.”
Privacy mode labels are marketing until proven otherwise. Local, Strict, Standard. These are Perplexity’s own definitions. There is no independent certification, no third-party audit of the data pipeline. Until that exists, treat privacy claims as aspirational.
Should you use Comet?
Try Comet if you do heavy research that involves comparing information across multiple sources, already use Perplexity for search and want tighter integration, or are willing to tolerate rough edges in exchange for early access to genuinely new browsing paradigms. It also helps if you spend significant time summarizing articles or extracting structured data from pages.
Skip Comet if you need a browser that never fails at basic tasks, rely on a specific extension ecosystem that may not work in Chromium forks, handle sensitive financial or legal data through browser automation, or value stability and predictability over novelty.
If you are somewhere in between, you have options. Atlas is worth comparing if you care more about reasoning depth than agentic actions. If privacy matters, Brave and Arc both take a local-first approach that is more verifiable than Comet’s self-defined tiers. And Comet is developing fast. The product in December 2026 may look different from the one you download today.
The bottom line
Comet is the most coherent attempt yet to build a browser where AI is the interface, not an add-on. Summarizing a dense article, organizing 40 tabs, extracting structured comparisons from product pages — when these work, it is hard to go back to a regular browser.
But the gaps are still big. It orders the wrong pizza. It forgets what you just did. It misses the interesting part of a podcast because everything looks equally important to a language model.
That does not mean you should not try it. If you spend your day doing research and are willing to double-check the output, Comet already earns its place. If you would rather wait until the rough edges are sanded down, that is reasonable too. The product is moving fast enough that checking back later is not a cop-out. It is just smart.