Perplexity Computer: New Multi-Model Agent Tool

Perplexity launches Perplexity Computer for advanced agentic workflows

Starting this week, Perplexity subscribers gain access to a new agentic tool: Perplexity Computer. The company describes it as a unified system that consolidates multiple AI capabilities into a single workflow engine. In short, the tool can run multi-step tasks autonomously using a mix of models and can spawn subagents to tackle specific problems.

How Perplexity Computer works

Perplexity says the Computer can orchestrate complex workflows across 19 different AI models. It runs entirely in the cloud and executes tasks such as gathering statistics, pulling legal or financial data, performing analysis, and producing finished outputs like websites or visualizations. The company's website shows example workflows that combine data collection, analysis, and presentation without manual intervention.

Perplexity Computer "unifies every current AI capability into a single system," allowing independent execution of complex workflows.

Availability and pricing

The tool is available now but only to Perplexity Max subscribers at $200 per month. Because it runs in the cloud, Perplexity says it may avoid some security risks associated with other agentic tools that run on local devices.

Product rollout and early issues

Last week Perplexity invited the press to a background briefing to preview the product and its roadmap. The company planned a live demo, but canceled it after discovering flaws hours before the event. TechCrunch has not conducted an independent hands-on demo yet.

Strategic shift: boutique users, enterprise focus

Perplexity appears to be repositioning itself. Once focused on broad consumer adoption with a search-like answer service and the Comet browser, the company now says it wants to serve a smaller, high-value audience — "people making GDP-moving decisions." Executives at the briefing (who requested anonymity) said they will prioritize enterprise subscriptions and deep research users over chasing mass monthly active user counts.

• Perplexity emphasizes revenue and enterprise metrics over MAUs.
• Executives stated they aim to deliver tools for deep research and high-stakes decision-making.

Benchmarks and claims

The company recently published Draco, a benchmark it built for complex research tasks. Perplexity says its deep research offering outperforms competitors like Google's Gemini on that benchmark.

Technology approach: multi-model orchestration and search API

Perplexity says it no longer relies on other companies' APIs for its web index and now operates an AI-optimized search API. The company still packages frontier models inside consumer-friendly interfaces and argues that orchestrating multiple third-party LLMs delivers the most cost-effective and accurate results.

"Multi-model is the future," a Perplexity executive said, noting that models are specializing rather than commoditizing.

Perplexity reports that users frequently switch models depending on the task: visual outputs were most often routed to Gemini Flash in December 2025, software engineering queries to Claude Sonnet 4.5, and medical research to GPT-5.1. The platform can automatically choose the best model for coding, copywriting, or research, or run multiple models at once through a feature called Model Council.

Transparency and past criticism

The company previously received criticism for running modified Chinese-built open-source models without clear disclosure. Executives now say transparent use of cheaper internal models can be an efficient way to optimize costs and performance.

Unit economics and product limits

Offering multiple models at flat subscription rates creates unclear unit economics. Perplexity claims it avoids huge infrastructure costs and that subscriptions yield high margins, allowing the platform to allocate tokens to the most suitable model for each task.

Roadmap and ecosystem moves

• Comet browser for iOS arrives next month.
• Perplexity will host a developer conference, Ask, on March 11 in San Francisco to encourage third-party API use.

Executives said they are now tracking recent revenue metrics more closely than daily query counts. Some community members have complained about new rate limits on free and paid tiers — a topic Perplexity executives dismissed at the briefing as inaccurate. One executive said claims that the free tier was being degraded or rate-limited are "completely false."

Where Perplexity sits in the market

Perplexity built early momentum by wrapping frontier models in simple interfaces and launching the Comet browser. Its total user base, estimated in the tens of millions, remains far smaller than competitors such as OpenAI, which reports hundreds of millions of users and has begun testing ads in ChatGPT. Perplexity previously experimented with advertising but abandoned it last year, saying ads undermined user trust in answer accuracy.

Perplexity's renewed focus on enterprise research, multi-model orchestration, and an AI-optimized search API positions the company as a specialist provider rather than a mass-market platform. For audiences tracking gaming news or other fast-moving fields, Perplexity's model-routing approach could help researchers combine specialized models to synthesize insights quickly and accurately.

Key takeaways

1. Perplexity Computer centralizes multi-model agentic workflows with subagent capabilities.
2. It's available only to Perplexity Max subscribers at $200/month and runs in the cloud.
3. The company is shifting toward enterprise, high-value users and deep research tools.
4. Perplexity emphasizes multi-model orchestration and owns an AI-optimized search API.
5. Benchmarks like Draco and features such as Model Council showcase the company's research focus.

As the AI ecosystem evolves, Perplexity appears to be betting on specialized, high-value workflows rather than broad consumer scale. Its new Computer tool and multi-model strategy aim to give enterprises and researchers a flexible, cost-efficient way to get accurate results from the best model for each job.