Perplexity Just Paywalled the Feature That Sold People on Its Browser. Budget for This Happening to Your Stack Too.
As of September 15, Perplexity moved Comet's "Control Browser" feature, the one that let the AI agent click, fill, and navigate on your behalf, out from under a standard Pro subscription and behind separate Perplexity Computer credits. It was one of the biggest reasons people gave for switching browsers in the first place. Paying subscribers are the ones who lost it, not people trying Comet for free.
What changed
Comet launched with agentic browsing as its headline feature: an AI that could take actions inside the browser on your behalf, not just answer questions about a page. Control Browser was the specific capability that made that real, and it shipped included with a Perplexity Pro subscription. As of this week, that's no longer true. The feature still exists, but using it now draws down a separate pool of Computer credits rather than being part of what your subscription already covers.
The numbers make this concrete. A fresh Pro subscription includes a one-time bundle of 4,000 Computer credits, and light Control Browser use can burn through that in a matter of days. After that, a Pro subscription doesn't refill the pool; you're buying more credits separately or going without. Perplexity Max subscribers get a one-time 35,000-credit bonus plus 10,000 credits every month, which makes the gap between what Pro and Max actually get you far wider than the headline price difference between the two plans suggests.
This lands on the exact group least likely to shrug it off: people who were already paying, who built the feature into an actual workflow, and who now find that workflow metered in a way it wasn't when they signed up.
The context that makes this land differently
This isn't happening in isolation. Perplexity has spent 2026 aggressively expanding Comet's reach: full iOS rollout alongside the existing Android, Mac, and Windows availability, Voice Mode moved onto OpenAI's GPT Realtime 1.5 with a reported 25%-plus improvement in interaction reliability, and Comet's agentic capabilities getting integrated directly into Samsung Internet as an optional default search engine on Samsung devices. By any distribution metric, Comet is growing, not struggling.
That's what makes the timing of the paywall notable. This isn't a company pulling back a feature because adoption disappointed them. It's a company whose distribution numbers are climbing, narrowing what a subscription includes at the same time. The two things aren't contradictory: more users means more compute cost for a feature that has an agent operating a browser session, and metering the expensive part while growing the free-to-try funnel is a coherent business decision. It's just not the decision the people who subscribed expecting Control Browser to stay included signed up for.
Why this matters beyond Perplexity specifically
If any part of your workflow depends on an AI vendor's "included with subscription" feature, especially one that involves the model actually taking actions rather than just generating text, this is the pattern to watch for. Agentic features cost meaningfully more to run than a chat response: they involve multiple model calls, tool execution, and often real browser or sandbox sessions that stay alive for the duration of a task. Vendors price these features aggressively at launch to drive adoption and prove the capability works, then adjust pricing once real usage data tells them what it actually costs to run at scale.
I watched this exact pattern play out with a different vendor's agent-execution API earlier this year: generous free tier at launch, months of building workflows around it, then a usage-based pricing tier introduced once the vendor had enough data to know the free tier was unsustainable at the volume power users were generating. The lesson wasn't "don't use the feature." It was "don't build a dependency on an included-in-subscription price for a capability that's expensive to run, without a plan for what happens when the vendor discovers the real unit economics."
What I'd actually do
If you or your product currently leans on Comet's Control Browser as part of an actual workflow, check your Computer credit balance and usage rate this week, not after you hit a wall mid-task. If the new metered cost changes your math, that's useful information now, while you have time to adjust, rather than during a client deliverable.
More broadly: treat any agentic AI feature (anything where the model is executing actions, not just generating text) as a cost center that will get repriced once the vendor has real usage data, regardless of what it costs today. Build workflows that degrade gracefully if that feature gets metered or removed, rather than ones that break entirely. That might mean keeping a manual fallback for anything Control Browser currently automates, or simply not routing anything client-critical through a feature you don't control the pricing of.
The honest take
I don't think Perplexity did anything predatory here. Running an agent that controls a browser session is genuinely expensive compute, and "included free during the growth phase, metered once the real cost is known" is a completely normal SaaS pricing arc, not a bait-and-switch. Where I'd push back is on how these changes get communicated: subscribers who built a real workflow around a specific feature deserve more warning than a quiet policy update discovered through outside reporting.
Where I could be wrong: it's possible Perplexity's Computer credits turn out cheap enough in practice that this barely registers for typical usage, and the backlash is louder than the actual financial impact on most users. I don't have per-task credit pricing to confirm either way as of this writing. What I'm confident about regardless: if you're building anything that depends on another company's agentic feature staying free, you're borrowing against a subsidy that has an expiration date you don't get to see in advance.
Author
Lukas
@lukcombinator