Confusion Takes Deep Research into Computing, Sub-Navigation Functions in 20+ Previous Models of Reports, Decks, and Dashboards

Confusion led Deep Research to Computer, its multi-model orchestration program. Improvements improve accuracy, depth of analysis, and quality of citations. Deep Research now breaks down complex questions into smaller tasks and guides them across 20+ parameter models. It returns actionable reports, decks, and dashboards, all within the desktop.

Advanced Research in Computing

Deep Research is a mode that uses multiple searches, reads sources, and writes a cited report. The new version resides within the Perplexity Computer, which launched in late February 2026. The computer is a cloud system that integrates up to 20 AI models into a single workflow. It is model-agnostic, with Opus 4.6 as its main logic engine. Sub-agents handle specialized work, such as Gemini for intensive research work.

Deep Computing Research is built in two parts: Agent Search SDK and Search as Code. With one complex question, create a research plan automatically. It then finds primary sources across hundreds of sites and cites every claim.

Search As Code: How It Works

The model writes the code that compiles the search itself. That code uses thousands of retrieval steps in parallel, directed at each query. The script runs in a sandbox and calls Perplexity’s Agetic Search SDK. The SDK introduces basic search functions such as filtering, recursive, and repositioning. This differs from a static pipeline that uses the same steps over and over again. Code-driven search allows the system to branch, compare, and refine as it learns.

Search as Code comes from both the desktop and the agent API. So developers can access the same search stack programmatically. The computer also reads your files alongside the live web. You can download a PDF or spreadsheet for the internal context. It then includes cross-references with census data, Statista, and other sources.

Example of a Working Engineer

Deep Research in Computer is a consumer feature for Perplexity Max users. Developers access the same stack through a pay-as-you-go agent API. The official SDK ships with a deep-research preset, shown below.

# pip install perplexityai
# export PERPLEXITY_API_KEY="your_api_key_here"
from perplexity import Perplexity

client = Perplexity()  # reads PERPLEXITY_API_KEY from the environment

response = client.responses.create(
    preset="deep-research",  # pre-configured research setup; "pro-search" is another
    input="Compare the cash flow and profit margins of the largest AI chip makers.",
)

print(response.output_text)  # aggregated report text from the run

The conclusion is POST https://api.perplexity.ai/v1/agent. It also agrees POST /v1/responses with OpenAI SDK compatibility.

Benchmark

Perplexity published before and after numbers comparing Legacy Deep Research to the PC version. The benefits are greatest in agent browsing, where the system must navigate through multiple pages.

BrowseComp tests an agent’s ability to find information that is hard to find by browsing. The jump from 40.7% to 83.8% is the biggest gain shown. The Humanities Final Examination includes expert questions on many academic subjects. DeepSearchQA is already sitting at the top, so its gain is small but good.

Use Cases, and Examples

Confusion sends the first tasks that indicate the target range.

• Finance: compare the cash flow and profit margins of the major AI chip companies over five years.
• Official: map how US and European data privacy laws differ in one comparison table.
• Health care: compile evidence from clinical trials on whether weight loss drugs improve heart health.
• Technology: the leading benchmark for models in imaging capability, cost, and context length.

Each job ends with a delivery. You can turn the report into a summary, deck, or live spreadsheet. The computer reads and writes from within the file, not to the side of it. It shows a preview before any change comes, approving or rejecting it.

Method Selects Models

The computer moves each subtask to the model that best suits it. A formal conceptual model handles contract review. The data model handles the variability of the spreadsheet. A writing model handles the final draft. Premium data sources return answers, including PitchBook and CB Insights. Legal data is currently being viewed first.

Powers and Limitations

Power:

• Code-driven search uses thousands of retrieval steps for each query.
• The biggest benefits are limited to agent browsing, led by the BrowseComp result.
• It reads internal files and the live web, quoting every claim in line.
• Produces deliverables ready for delivery: reports, summaries, decks, dashboards, and live spreadsheets.

Limitations:

• The standing numbers are preliminary, so independent verification is still important.
• The In-Computer feature is focused on Perplexity Max, not the free tier.
• Premium source coverage varies, and legal data is always in preview.
• The output still needs human review, because “quoted” is not always correct.

Key Takeaways

• Confusion has moved Deep Research into Computing, directing research sub-tasks across 20+ frontier models.
• “Search As Code” allows the modeler to write code that uses thousands of retrieval steps in parallel.
• BrowseComp accuracy jumped from 40.7% to 83.8%; The Final Humanity Test increased by 36.4% to 50.5%.
• It reads your files and the live web, citing all claims in all reports, decks, and dashboards.
• Developers can access the same agent search stack through a pay-as-you-go agent API.

Check it out Technical details. Also, feel free to follow us Twitter and don’t forget to join our 150k+ ML SubReddit and Subscribe to Our newspaper. Wait! are you on telegram? now you can join us on telegram too.

Need to work with us on developing your GitHub Repo OR Hug Face Page OR Product Release OR Webinar etc.? contact us