Last week, M.G. Siegler at Spyglass officially declared that the browser wars had begun. Literally one week later, it's already a three-way cage match.
On the same day, Reuters dropped a bombshell about OpenAI's entrance to the race at the same time Perplexity launched its Perplexity Comet browser. Meanwhile, the upstart browser Dia has been out in beta for a few weeks now.
Three AI browsers are now battling for your attention:
- OpenAI just had plans leaked for an AI browser launching in a few weeks.
- Perplexity Comet started rolling out to $200 Max Tier subscribers.
- Dia is already live in beta for Arc users
The timing isn't coincidental. Each represents a different vision for AI-powered browsing.
Let's break down what each of these AI browsers actually does:
Dia is an interesting case; let's start there, since it was the first.
Founded by The Browser Company, Dia is the next evolution of Arc Max, which is sort of the O.G. AI browser (if such a thing could be said).
It's already available in beta to Arc users, making it the most mature of the three.
Why is it the "next evolution" of Arc? Because The Browser Company made the bold decision to essentially put Arc in maintenance mode and go all-in on this AI-first browser.
The content platform / AI product creator Every had a great interview with Josh Miller and Hursh Agrawal of The Browser Company about Dia, and how they had to kill Arc to bring Dia to life.
Why kill a beloved product? CEO Josh Miller was brutally honest: "For most people, Arc was simply too different, with too many new things to learn, for too little reward." Only 5.52% of daily users used more than one "Space" feature, and just 4.17% used Live Folders.
What makes Dia different:
It lets you chat directly with your tabs. Instead of copy-pasting links into ChatGPT, the AI sidebar can analyze all your open tabs at once and answer questions about their content.
Real-world examples:
- One user said "Dia cuts the busywork. With just a prompt and tabs, I find sources in minutes, not hours."
- Another uses it as an "eng tutor" that "answers questions in-line and gives me practice problems."
- Content creators use it to "stay pitch perfect across all our client brands" with personalization features.
The browser looks like Chrome (which people actually want), but with AI baked into every interaction. It can email participants separately from a meeting, suggest writing improvements, and even help with shopping decisions by comparing options with context.
The memory feature: Dia can learn from your past seven days of browsing history to provide personalized recommendations, whether you're shopping or need writing that matches your style.
But Now, the Perplexity Comet Has Arrived
But the big launch today is Perplexity Comet. That’s AI search engine Perplexity’s “agentic” web browser.
Perplexity describes it as a browser for "agentic search," which means AI that can perform multi-step tasks autonomously. Think less "answer my question" and more "book this reservation for me."
The marketing promises are bold: "Comet makes Perplexity available on every site you visit. Quickly understand anything you see, in any language, or from any perspective."
Planned features:
- Comet will be free for all users and support importing all your Chrome extensions and bookmarks.
- Uses Perplexity's search engine instead of Google.
- CEO Aravind Srinivas demonstrated natural language commands like "reopen the recipe I was viewing yesterday."
The privacy angle: Three privacy modes: no memory/no ads, memory without ads, or memory with ads, with data stored locally and not used for AI training.
Check out the video above for Matt Berman’s hands-on demo with Perplexity Comet, or if you’re already a Max subscriber, download it here. He explains how it works with an AI agent alongside your search engine, which is pretty interesting, so we'll share a few of his observations:
First of all, Matt says Comet is a Chrome fork that imported all his settings, bookmarks, and extensions automatically. He noted it's "lightning fast" - even faster than regular Chrome. He walks through a few tests he actually did with Perplexity, and how they went:
- LinkedIn automation: "Find interesting contacts in my connection requests" → "Accept all of those requests" (worked perfectly).
- Shopping research: Asked it to check multiple electronics stores for Nintendo Switch 2 inventory (worked after being explicit: "go check for me").
- Calendar integration: "Who am I meeting today?" pulled from Google Calendar.
- Twitter posting: Successfully found and quote-tweeted his recent post.
Crucially, he also shared a few tests that DIDN'T go his way:
- YouTube replies: Found the top comment but couldn't auto-reply due to "technical restrictions."
- Unwanted branding: Added "created with comet assistant" to his tweet without permission.
- Needs explicit instructions: Sometimes just gave suggestions instead of taking action.
Matt sees a future where you interact with agents instead of websites directly. Multiple agents running in parallel handle tedious tasks (flight booking, grocery shopping) while you just approve the final results.
His take: "I don't want to go through the process of booking a flight. I just want it all done for me and then at the last moment, let me just approve it."
The browser control feature shows blue outlines when the agent is working and hands control back when finished - though this was a bit buggy in his demo.
Now, why would Perplexity build a browser? As a user, why would you use an AI browser vs. just use Perplexity?
Matt figured out three key reasons:
- First, platform risk: You can avoid relying on Google/Apple browsers that will prioritize their own AI.
- Second, no re-authentication: Today's cloud-based browser agents (like Operator, or any number of computer use agents) require logging into everything again; Comet uses your existing logins.
- Third, context awareness: You can pick up from where you already are on a website instead of starting over.
Which brings us to OpenAI...
OpenAI's Mystery Browser: Coming in Weeks
This morning's Reuters exclusive revealed that OpenAI is "close to releasing an AI-powered web browser that will challenge Google Chrome" in the "coming weeks".
Here's what makes it interesting:
The browser is designed to "keep some user interactions within a ChatGPT-like native chat interface instead of clicking through to websites". That's a fundamentally different approach—rather than enhancing web browsing, they're trying to replace it wherever they can (we'll get into why below).
Here are the tech specs:
- Built on Chromium (like most browsers).
- Will integrate Operator, OpenAI's web-browsing AI agent.
- Enables AI agents to "carry out tasks on behalf of the user, like booking reservations or filling out forms."
The strategic play is obvious: If adopted by ChatGPT's now 400-800 (we've heard both) million weekly users, it "could put pressure on a key component of rival Google's ad-money spigot".
Why now? OpenAI hired two longtime Google VPs who were part of the original Chrome team, and they've made it clear they'd want to buy Chrome if antitrust enforcers force a sale.
Since OpenAI’s browser will work similarly to Perplexity’s, incorporating OpenAI’s Operator “computer use” agent, Matt Berman points out this tactic gives both companies access to a lot more data to work with, and also gives them protection from Google turning Chrome’s native AI experience into the dominant way users interact with AI.
Example: if Google were to turn Chrome into its own Gemini-first browser, and it was actually good, then you’d be using Gemini everywhere instead of going to another website like ChatGPT or Perplexity.
This is all bad news for publishers, who are already seeing their traffic and revenue decline rapidly just from “AI Overviews” they are pursuing antitrust action against Google. https://support.google.com/admanager/answer/13860694?hl=en
And as The Information writes, they might have a pretty good case; since publishers rely on Google for traffic, they can’t block its AI crawler (since its the same as their web index crawler). This gives Google all of the leverage, and publishers no real way to opt out.
Google already had an antitrust ruling against them back in April in the US, and any number of options are on the table to force them to give up some of their monopoly power. One option is forcing a sale of Chrome (which would benefit OpenAI), but another could be resolving this whole problem of crushing publishers.
Allow a tangent on the new business model of Web 3.0:
Last week on the More or Less podcast, the gang debated the case of Google versus the publishers.
The crux of the discussions was this:
- First, Jessica Lessin (founder and CEO of The Information) noted, "We're in a post-traffic world, you guys." Publishers can't rely on Google or social traffic anymore... the only option is to "feed everything to the AIs now" or hide behind paywalls.
- The hosts absolutely roasted Google's new "Offer Wall" product—a desperate attempt to help publishers monetize by letting readers take surveys or watch ads for micropayments. As they put it, whoever's PMing this at Google is "just thankful they still have a job."
- Sam hilariously suggested Google should just give one random publisher $100 million annually as a "Publishers Clearing House"—complete with a giant check at the door. It's absurd, but captures the desperation of the situation.
- And they basically laughed-off micropayments.
The strongest argument from the podcast? That we're seeing the last stand of Web 2.0 (shout out Dave Morin for that angle; I love it!) If that's true, then "traditional browsers themselves" might be the real casualty of this war. Which makes the current browser war feel less like competition and more like a farewell tour to Web 2.
Why do we bring this up? Because AI search begs for a different business model
Despite the More or Less fam's dismissal, we still think the only economic model that makes sense IS micropayments to publishers (no matter what the Lessins think). In my mind, it would go like this: every time an AI uses a publisher's information in a user query, they get paid. You could go so far to charge per crawl, like Cloudflare is proposing. Either way, you can think of these payments like cost per token (how AI companies charge API users of their services), but in reverse, where each token from a publisher costs x cents per million. It might drive up AI costs, but it’s the only thing that’s fair if you’re using their content, right?
Yeah, well, here's the problem with that (according to the More or Less gang, as much as WE like the idea):
- The hosts noted that AI companies default to pay-per-token models not because it's ideal, but because they literally can't afford unlimited plans. "Super users use so much, they just kill you."
- This mirrors early fintech where expensive human-in-the-loop processes forced usage-based pricing.
- Regardless of whether or not publishers would actually make meaningful revenue from micropayments, you add another cost center to AI companies who are already bleeding loss leaders, annnnd that's not gonna be an economically sound system, any way you slice it.
Now, the More or Less crew did say the ideal model would charge based on value created—like hedge funds taking a percentage of returns. But as they pointed out, "No CFO will do that deal." Companies culturally resist paying tech what it's worth, even if it's a float of margin, because (if I'm interpretting this correctly) they want fixed costs to anchor to even if a company could say "I'll increase your production by 100%, and you just give me 50% of the return."
So the Morins and the Lessins are right that usage-based models emerge from weakness, not strength. But IMO, they're thinking about it backwards—it's not about charging users per token, it's about AIs paying publishers per token consumed. In a way, that's the "pay per value" model in action. Only when a user gets information that's actually USEFUL to them from a publisher do the publishers get paid.
How would you track something like that? Easy, the AI companies already do it through reinforcement learning with human feedback. You ask the user: "was this information helpful?" or better yet "was this the information you needed?" And they give a thumbs up or thumbs down. Then, when the user gives a thumbs up, the money for that request goes to the publisher, rewarding publishers for writing and sharing valuable content.
In theory, it incentives the publishers to produce actionable, useful content without filling pages upon pages of words no one actually wants to read to get the information they actually want, just to sell ad space.
Users, by extension, will have a wealth of knowledge about themselves stored up in these agentic browsers. I mean really, if you go far enough down this road, then individuals should be able to sell their data on a per token level just like publishers could.
This data too could be consumable by advertisers. Say an advertiser wants to sell you something. Why couldn't they request a certain fee to you, through your browser, to get some information about what you like or dislike?
This brings us closer to Tim Berners Lee idea of users controlling their own data in data pods. Your AI account, or if incorporated in a more intertwined way, your AI Browser will be like your own personal data pod. And if you have your own custom agent (or knowledge base that other agents can access), you could charge for access to that data and get paid for your information (maybe in some sort of query by query system, where you can approve information before its shared).
Back to publishing: this is a problem I think about a lot, and it's not an easy fix.
Publishers have a hefty burden; everyone benefits from the value of information shared freely online, but only one person (the publisher) shares the burden of cost. Sam Lessin talks about this problem a lot too, and how the really valuable information ultimatley is gatekept by paywalls or being in the room with the handful of people who can tell you what's really up (and that's the only way it ever works; I philosophically want him to be wrong, and that a better way can and should exist, but realistically think he's probably right).
Okay, fine, so apply the same concepts here: perhaps some kinds of information is actually really valuable, and it's worth paying for. Just like how access to o3 is more expensive on a cost per token basis. Maybe THAT information is charged at a higher token per consumed basis, too.
Compare this to the current subscription system: maybe it's not worth it for everyone to pay expensive subscriptions for data they only need access to one time. But maybe it's worth A LOT for them to pay for it one time.
So maybe you paywall specific queries and/or data at a higher tier. And if the user wants to access, say, The Information's data center buildout database, they could pay the AI search engine (and by extension, The Information), a higher one time consumable fee for those tokens, because it's worth it for them to get that data for whatever project they're working on.
Think about it like this: The browser wars make resolving this issue urgent. If OpenAI succeeds in keeping users inside ChatGPT, publishers won't even get the courtesy of an every now and then click.
People are using AI search engines, for better or worse, because they are a better experience. Just the information you want, none of the information you don't. In theory, not in practice, because there's still plenty of halluciantions. But if you know how to work with these tools and know what "correct" looks like when you see it, you get incredible upside from using AI search. It really can be a better user experience, and at the end of the day, capitalism is going to reward better user experience (or encumbent superpower, and right now, the encumbent superpowers are using said super powers whether publishers like it or not).
Now, is all this wish fulfillment on how things should and could be easier for the user? Yes of course. Could this also be a bad business model for everyone involved? I acknowledge that I'm likely wrong, and that yes, it could still be a bad idea.
So let's say then no one in the market would go for a pay per token consumed model. Maybe that's not valuable enough, or too scary for folks, or crushes one side or the other and isn't a functional market. To me, it feels like the perfect use case for blockchain and/or crypto, even if that's not how its applied (which Sam loves), but okay, fine, I'll accept that no one wants this.
Alternatively, AI companies (Google especially) should just steal YouTube's model and apply it to AI search:
If micropayments won’t work, then apply the YouTube business model to search: show contextual ads alongside AI responses (real offers with dynamic discounts, not garbage banners), let users downvote brands they hate, and split revenue with the content sources. Publishers get 50% if users thumbs-up the answer, 10% if they get a thumbs down (or nothing), 30% baseline if the user does nothing.
Think about it: every time you repurpose content from a publisher, you can show ads on the side. We all hate ads as consumers, but need them as publishers, so AI companies will need to make the ads they provide actually good, and by good, I mean useful for the searcher.
We're not talking about banner ads here, but like competitive offers from rival services related to what the user is searching for. Maybe these offers could include real-time generated discounts.
Also, you heard me: this might be my hobbyhorse, but it's time we let users actually DOWN VOTE ads. Let us shoot down the offers from brands we don’t want. Since AI companies already use reinforcement learning with thumbs up and down, users should be able to downvote ads they don't want. Then users will stop seeing ads from companies they’ll never buy from, and advertisers will stop paying to reach users that’ll never become customers.
This is my big personal gripe with YouTube ads as they stand now. There’s no way to say “stop showing me this brand please God” without accusing them of spam or immoral content. Which btw, if I have to see another Youtube ad for “Motion AI”, "Grammarly" or "Monday.com" I might just click that "this ad is inappropriate" button...
Why can't you downvote ads today? I mean really, is it not the best form of feedback anyone could give? "I don't want this. Stop wasting your money trying to sell me." Great, move on. Offer me something else, or offer your product to someone else who does want it.
I get that in a previous world, everyone would down vote every ad they ever see, but with "agentic browsers", you'll be able to give so much more concrete feedback on what you want and why (and have your browser, and by extension, advertisers remember that) that it'll be worth it for both sides.
However you do the ad portion, the main idea here it to share a meaningful portion of that ad revenue with the originators of the content that gave the user the answer they were looking for. You could gate the pricing, say the publishers gets 50% rev-share if the user votes a thumbs up that the information answered their question, or a flat 30% if no feedback is given. Something that incentivizes higher value content AND a healthy info-sharing ecosystem. It won't be perfect, but AI allows us new ways of imagining advertising and digital communications. Let's use that power in a way that benefits the entire ecosystem.
The advent of AI Browses may very well be the catalyst that kicks this off.
Why AI Browsers Matters for Your Daily Workflow
These browsers represent three different visions for how AI should integrate into our digital lives:
- Dia enhances traditional browsing with conversational AI.
- Comet promises autonomous task execution while you browse.
OpenAI's browser wants to eliminate browsing entirely for many tasks.
The winner will likely be whoever figures out the right balance between convenience and user control. Early Dia users report loving the seamless integration, but some worry about data privacy. Perplexity's privacy-first approach with Comet could be appealing, but we'll see if it sticks.
The real disruption: All three browsers are trying to solve the same fundamental problem—the web wasn't designed for AI, so they're building new interfaces that are.
Our prediction: As M.G. Siegler noted, "traditional browsers, as we know them, will die. Much in the same way that search engines and IDEs are being reimagined." Couldn't have said it better ourselves! Who likes actually having to type and search as much as we do? A much more collaborative, fluid system makes much more sense.
So yeah, traditional browsers will die, slowly then suddenly. And when they do, the entire economic model of the internet goes with them. Publishers, advertisers, and users better figure out what comes next—because "next" is apparently just a few weeks away.
