Summary: OpenAI updated Deep Research on February 10, 2026 so it can connect to any MCP server or app and restrict web search to trusted sites. The update also moved the feature to GPT-5.2 from the earlier o3 and o4-mini models. Plus and Pro received it that day, with Free and Go following within days.
February 10, 2026 is the day OpenAI rolled out a sharply redefined Deep Research inside ChatGPT. The update note on the official page wastes no words: deep research can now connect to “any MCP or app,” and web search can be confined to trusted sites, keeping research anchored to authenticated, industry-standard sources. The Decoder filled in the other half of the story that day: the feature now runs on the new GPT-5.2, taking over from the o3 and o4-mini models that had carried it since the 2025 launch.
If you write software or work with information for a living, there is more here than a fresh coat of UI paint. Where a research agent draws its data is moving from “the open web” to “the connectors and sites you designate,” and MCP now officially serves as an input surface for a capability built into ChatGPT itself.
What Shipped, Concretely
Put the official update note next to the ChatGPT release notes and four concrete changes emerge:
- Any MCP server or app is fair game as a connection, letting Deep Research pull data from tools and services you have already authorized
- Web search can be fenced into specific websites, shrinking the source set down to domains you trust
- Runs are tracked live, and you can step in mid-run with follow-up prompts, additional sources, or a change of direction
- The sidebar entry point and fullscreen report view got a redesign, and research plans can be written and edited before a run kicks off
The release notes spell out the rollout as well: Plus and Pro accounts received it over the course of the day, with Free and Go accounts to follow within days.
The Real Story Is Trusted Sources
Deep Research’s core promise has always run the same way: fan out across many steps and a large volume of pages, then synthesize the findings into a long report. Open-web search fails structurally in two places, though — you cannot control source quality, and anything behind a login stays invisible.
Fencing search into specific sites puts the corpus boundary under the user’s control — a legal team can limit queries to regulator domains, an engineer to official docs and issue trackers. MCP apps close the other half of the gap: information living in company wikis, databases, and internal tools can now flow straight into the research process. The release notes name the objective without adornment: producing “more accurate, credible reports with greater control.”
One caveat from The Decoder belongs in your notes: web access lowers hallucination rates without driving them to zero, and risk piles up as reports get longer. Better-constrained sources mean better inputs — not a guarantee of better outputs.
From Community Protocol to Product Default
The Model Context Protocol arrived in November 2024, introduced by Anthropic. When OpenAI adopted it in March 2025, most observers still shelved it under developer-ecosystem concerns. This update is where that changes: MCP has outgrown its role as an API-level connector option and become the standard input for ChatGPT’s flagship research capability.
For the MCP ecosystem, the effects split in two. For one, every service exposing a remote MCP server suddenly serves a new class of end users — researchers rather than developers. For another, Deep Research itself is being recast: once a report generator aimed at the public web, now an engine that synthesizes your authorized data and designated sources. Nor is this an isolated move — the lightweight version arrived in April 2025 and the product line was folded into Agent mode in July 2025 — and this update extends the same trajectory: the research agent is settling in as a resident mode of ChatGPT rather than a standalone feature.
Where This Lands in Real Workflows
Three things stand out. One: with editable plans and mid-run steering, human-in-the-loop control now sits inside the agent’s execution loop — a course correction at minute five beats waiting twenty minutes for a report that drifted off target. Two: source restriction opens compliance- and legal-sensitive work to Deep Research for the first time — teams that could never risk an agent roaming the open web. Three: the trust boundary around connectors turns into a genuine security duty, because somebody has to audit what each connected MCP server can read.
Elsewhere on the same day, OpenAI quietly refreshed the response style of GPT-5.2 Instant; both entries share the February 10 release notes. A base model and a research agent shipping in the same batch is a cadence worth watching.