Wednesday Jun 17
PRODUCTOther

Perplexity Splits Research Across 20 Models

17JUN
20+ MODELSPERPLEXITYONE REPORT

Perplexity thinks no single model should answer hard questions. Its Deep Research now splits a query into subtasks across 20+ frontier models. It returns one report, deck, or dashboard.

The app layer is going model-agnostic. The winner picks the best model per subtask, not one model for everything.

Deep Research runs inside Computer, Perplexity's orchestration system. It breaks a question into parts and sends each to the strongest model, from Claude to Gemini to GPT. The output is a finished deliverable, not a chat reply.

For PMs, this is the commoditization signal. If routing wins, model brand fades behind the workflow. The risk: orchestration quality is hard to verify and easy to fake.

full brief & sources

⚡ Why this matters

  • Multi-model routing treats frontier models as interchangeable parts.
  • Value moves from owning a model to orchestrating many of them well.
  • Knowledge-work output becomes the product, not the chat answer.

🔍 What happened

  • Jun 11: Perplexity moved Deep Research into Computer, its multi-model orchestrator.
  • Hard questions are split into subtasks routed across 20+ frontier models.
  • Models used include Claude Opus, Gemini, and GPT variants.
  • Output is work-ready: reports, slide decks, spreadsheets, and dashboards.
  • Perplexity claims better accuracy, depth, and citation quality.

💬 Smart takes

  • MarkTechPost: Deep Research now returns work-ready reports, decks, and dashboards.
  • OpenTools: Computer coordinates up to 20 models in one workflow, model-agnostic by design.
  • Skeptic: routing across 20 models multiplies cost and latency. Users may not see enough quality gain to justify it.

🧭 Where this goes

  1. Likelymore apps add model routing instead of betting on one lab.
  2. Likely'deliverable, not chat' becomes the default bar for research tools.
  3. Possiblelabs fight back by bundling orchestration so you stay in-house.
  4. Possibleenterprises demand an audit trail for which model did which subtask.
  5. Wild Cardmodel brand becomes invisible to most users within two years.

🥄 The Spoon Take

This is the commoditization bet in product form. If the best answer comes from routing across many models, no single lab owns the user. The app that orchestrates wins. Labs will resist by making their own model the easiest default. Watch whether routing quality gets good enough that people stop caring which model replied.

🤔 Pushback

Orchestration is invisible plumbing. If the routed answer is not clearly better than one strong model, users pick the simpler, cheaper single-model tool.

·MarkTechPost - Deep Research routing·OpenTools - Multi-model Computer·BuildFastWithAI - Perplexity Computer