Comet Browser by Perplexity AI vs Sup AI

“The category is agentic browser automation — direct competitors are Anthropic's Computer Use, OpenAI Operator, and Arc's now-shelved Browse for Me, all of which have demonstrated the same core loop and hit the same walls: form auth, CAPTCHAs, and any site that detects non-human behavior. Comet breaks the moment a user wants it to handle a logged-in, dynamic SPA that rate-limits bots — which is most of the web that matters. What kills this in 12 months: OpenAI ships Operator to all ChatGPT users for free and Perplexity's differentiation collapses to brand preference. To earn a ship, Comet needs to demonstrate persistent session handling and a credible story for the 60% of high-value tasks that live behind auth walls.”

“Extraordinary claims require extraordinary evidence. A 7.41 point jump on HLE via ensembling — without publishing methodology — smells like benchmark gaming. The latency of running 339 models in parallel is also a real concern for anything other than async research tasks.”

“The thesis here is specific and falsifiable: by 2027, the browser tab is no longer a viewport you stare at — it's a task queue you delegate to. Comet is betting that the interface layer between humans and the web collapses from 'navigate and click' to 'state intent and verify result.' That's a real trajectory, and Perplexity is one of the few players with a live search index plus the intent-capture surface to make the delegation model feel natural rather than scripted. The second-order effect that matters: if Comet works, SEO as a discipline dies faster than anyone is modeling — the bot reads the page so the human doesn't, and click-through becomes irrelevant. The dependency that has to hold: users must be willing to hand over ambient browsing context to Perplexity's servers, which is a trust bet that sits on regulatory quicksand. Still, as a positioned bet on the trend of intent-first computing, this is early and credible rather than late and derivative.”

“Model ensembling is an underexplored direction in the race to reduce hallucination. If Sup AI's approach scales, it could be more durable than fine-tuning individual models — you get the wisdom of the crowd across model families, training data, and architectures simultaneously.”

“The buyer is a Perplexity Pro subscriber who already pays $20/month — Comet is a retention and upgrade mechanism dressed as a product launch, which is actually smart distribution. The moat question is harder: browser distribution is a graveyard (ask Opera, Brave, Arc) and the switching cost of a browser is enormous for consumers but thin for Perplexity because users won't abandon Chrome for search features alone. The business survives model cost compression because Perplexity's value isn't the underlying LLM — it's the index and the task orchestration layer sitting on top of it. What worries me is the expand story: once you've automated the tasks a Pro user cares about, what's the upsell? There's no obvious enterprise tier with audit logs and admin controls mentioned at launch, which means the revenue ceiling is whatever the Pro subscriber count is. Viable, but not yet a standalone business thesis.”

“The job-to-be-done as stated is 'complete multi-step web tasks autonomously' — that sentence contains an 'and' hiding inside 'multi-step,' which means this product is trying to solve task delegation, context retention, and web navigation simultaneously before nailing any one of them. The onboarding reality: users join a waitlist, get access inside a Pro subscription, and then face the blank-slate problem of not knowing which tasks are reliably automatable versus which will silently fail halfway through. That's not a 2-minute path to value — that's a discovery tax. The product isn't complete enough to replace any existing workflow today because there's no task library, no failure transparency, and no way to audit what the agent actually did. Until Comet ships a defined set of tasks it handles end-to-end with high reliability and surfaces that clearly at onboarding, it's a demo with a waitlist, not a product.”

“The HLE claim needs independent verification, but the underlying ensemble approach is architecturally sound for factual Q&A tasks. Running 339 models is expensive — pricing will be the gating factor for production use. The $10 free credit is a fair trial.”

“For creative work, ensemble outputs tend to regress toward the mean — you get the most-agreed-upon version of something, which is usually the least interesting version. This is a tool for factual accuracy, not creativity. I'd stick with a single strong model for writing.”