Key insights

• Perplexity is positioning itself as an orchestration layer above the model providers, betting that routing tasks to the best model matters more than building one model to rule them all.
• The Amazon court ruling over Perplexity's Comet browser could set a precedent for how any AI agent is allowed to interact with websites on a user's behalf.
• If AI models keep specializing rather than converging, the company that picks the right model for each task gains an advantage that grows over time.

In Brief

Perplexity held its inaugural Developer Day and launched over a dozen new products, headlined by "Computer," a cloud-based AI agent that combines a browser, file system, and terminal into a single digital worker. CEO Aravind Srinivas told CNBC's Andrew Ross Sorkin that Computer orchestrates multiple AI models rather than relying on one, arguing that models are specializing rather than becoming interchangeable. The interview also addressed Amazon's court victory to block Perplexity's Comet browser from accessing its platform, a case with implications for how AI agents interact with the web. As AI companies race to build agents that can act on users' behalf, Perplexity is betting that the real value lies not in any single model but in knowing which one to use and when.

---

What Computer actually is

Most AI products give users one model and one interface. Perplexity's Computer takes a different approach. Srinivas broke it down in simple terms: think about what you need to do work on a computer. You need a browser to access the internet, a file system to store documents, and a terminal where programs can run. Computer puts all three inside a secure sandbox running on cloud servers, then pairs them with large language models (LLMs), AI systems trained on vast text data, that decide what actions to take.

The result is what Srinivas calls "a digital worker on the cloud" with access to a user's personal data and persistent storage, meaning files and preferences carry over between sessions. Unlike a ChatGPT conversation that starts fresh each time, Computer maintains context and can pick up where it left off.

Sorkin compared Computer to open-source projects like OpenClaw, noting it felt similar but with better security. Srinivas agreed that security is a key differentiator. Because Computer runs server-side in a sandboxed environment (an isolated space where code runs without affecting the main system), it cannot override user files or import malware. Perplexity also built its own Comet browser to handle agentic browsing, where the AI navigates websites on the user's behalf, rather than relying on third-party browsers.

---

The specialization bet

Sorkin raised a natural question: if every AI company builds something like Computer, won't the underlying models eventually become interchangeable? Srinivas pushed back firmly. "That's exactly what is not happening," he said.

His argument centers on a trend he says accelerated in late 2025. Rather than converging on the same capabilities, models started specializing in different strengths. Anthropic excels at coding. OpenAI's models are strongest in writing. Google's models lead in multimodal tasks like image and video understanding. Even within coding, different models handle different sub-tasks better.

Srinivas cited Anthropic CEO Dario Amodei making the same observation in recent interviews: that even within a single skill like coding, OpenAI and Anthropic models have different strengths. Perplexity's pitch to enterprise customers is that they don't have to pick one horse in the race. Instead, Perplexity figures out "which model is best for what purpose" and routes each task accordingly.

This is a fundamentally different business model from what OpenAI or Anthropic pursue. They sell their own models. Perplexity sells the routing. Srinivas called his company "the orchestra conductor across all these different models."

---

The Amazon court fight

The conversation shifted to a legal battle that could shape how all AI agents interact with the web. Amazon won a court order to block Perplexity's Comet browser from accessing its platform, arguing that the AI was scraping product data.

Srinivas framed the dispute with a store analogy. Imagine you walk into a shop and bring a friend to help you make a buying decision. The store owner says you can only use their staff for advice. "Who is actually right here?" he asked. Perplexity's position is that users should be allowed to bring their own AI assistant to any website.

He pointed to a detail in the preliminary injunction ruling (a court order that temporarily blocks an action while a case is decided). The judge reportedly noted that "Perplexity is actually useful to the user and is not doing anything egregious." Srinivas expressed confidence that Perplexity will ultimately win.

Sorkin pressed on a separate concern: even if Perplexity helps the consumer now, browsing on users' behalf could let the company collect data used in unintended ways later. Srinivas responded that Perplexity's browser is "first class in terms of how secure it is." He cited published evaluations on reliability and prompt injection resistance, where malicious instructions trick an AI into ignoring its rules.

The case matters beyond Perplexity and Amazon. As AI agents increasingly browse, shop, and negotiate on behalf of users, the legal question of whether "user permission" equals "platform authorization" will affect the entire industry.

---

Practical implications

For AI product builders

Perplexity's multi-model approach suggests that companies building AI agents may benefit from staying model-agnostic rather than locking into a single provider. If specialization continues, the ability to route tasks to the best model becomes a competitive advantage.

For enterprise buyers

The orchestration pitch is appealing but raises vendor-lock questions. Enterprises currently using a single AI provider might benefit from Perplexity's model-agnostic approach, but they're trading one dependency for another. The value depends on whether Perplexity's routing consistently picks better models than what a single provider offers.

For web platform operators

The Amazon case signals that websites may need to develop clearer policies on AI agent access. The distinction between a user accessing a site and an AI agent accessing it on the user's behalf is a question without settled law.

---

Glossary

---

Sources and resources