OpenAI is quietly transforming its ‘deep research’ agent from a broad web scraper into a precision tool for professionals who can’t afford hallucinations from random blogs. The February 10 update is the real signal here: you can now chain the agent to any MCP or app, and crucially, restrict its web searches to a curated list of authenticated, industry-standard sources. For an analyst at a hedge fund or an engineer sourcing technical specs, that’s the difference between a useful report and a liability. The company has been iterating fast since the initial February 2025 launch, first bringing it to Plus users, then giving it a visual browser in agent mode by July, and now layering in this enterprise-grade control.
Under the hood, the heavyweight version runs on a tuned-up o3 model that uses reasoning to pivot between searches, crunching through text, images, and PDFs. It’s designed to work for tens of minutes, doing what a human research analyst would need hours to complete. OpenAI frames this as a step toward AGI that produces novel scientific work, but the immediate use case is more grounded: giving knowledge workers a way to offload the grunt work of synthesizing hundreds of sources into a single, fully cited report. The output is methodical—a summary of its thinking plus clear citations for every claim.
Cost has been the obvious bottleneck, and the April 2025 update addressed that with a tiered system. Pro users get 250 queries per month on the full model. Plus, Team, Enterprise, and Edu users get 25. Once you hit that limit, you’re bumped to a ‘lightweight’ version powered by o4-mini, which OpenAI claims preserves quality while being cheaper to run. Free users get five queries, period. It’s a classic enterprise funnel: get the power users hooked on the good stuff, then meter access to manage compute costs.
The latest visual refresh and real-time progress tracking make the experience less of a black box. You see the steps it’s taking and the sources it’s pulling, and you can interrupt to refine with follow-up prompts or inject new sources. That interrupt capability is quietly significant—it turns deep research from a fire-and-forget query into something closer to an iterative conversation with a junior analyst. The real question now is whether restricting to trusted sources makes the agent more reliable or just more limited. For regulated industries, the trade-off is a no-brainer.
