Anthropic is pushing Claude further into long-form research work. The company says its upgraded Research feature can now spend up to 45 minutes on complex investigations before returning a cited report, while a new Integrations feature lets Claude search connected applications.
The pitch is straightforward: give Claude a difficult question, let it break the work into smaller pieces, and receive a document-style answer with links back to source material. The caution is just as important: AI research reports can still include plausible errors, mismatched citations, or quotes that need independent checking.
What changed in Claude Research
Anthropic announced the upgrades on Thursday. Claude already had a Research feature, announced on April 15, and the new version expands how deeply it can investigate a request.
When a user turns on the Research button, Claude can divide a complex prompt into subquestions, inspect each part, and assemble the results into a report. Anthropic says the upgraded mode can work across "hundreds of internal and external sources" and include citations that point back to original sources.
The longer runtime is the headline change. Anthropic says most reports finish within 5 to 15 minutes, but particularly complex investigations can now take up to 45 minutes. The company frames those longer sessions as a way to handle tasks that would typically require hours of manual research effort.
Claude is not alone in this category. Google’s Deep Research debuted on December 11, ChatGPT’s deep research features arrived on February 2, and open source clones of the approach have also appeared. Across these systems, the core idea is similar: autonomous browsing and source gathering, followed by a structured research document.
Why the extra time matters
A 45-minute research run changes expectations for an AI assistant. Instead of a quick answer, Claude Research is being positioned as a tool for gathering, organizing, and summarizing material that may be spread across many places.
That can be useful when the user does not already know where the relevant information is located. Deep research tools can surface sources that may be hard to find through ordinary search habits, especially when the question has several parts or requires comparing information across documents.
But more time spent searching does not automatically mean the final report is reliable. The source article notes that these AI research systems can produce confabulations: plausible-sounding material, including invented sources, inaccurate claims, or source references that do not support the text as written.
That makes the feature most valuable when the person reading the report can evaluate the subject. A knowledgeable user can often spot weak sources, irrelevant evidence, or claims that need confirmation. A user who accepts the report without checking may miss those problems.
A test shows both promise and risk
The source article describes a simple test using the question "Who Invented Video Games?" Claude Research spent 13 minutes and 2 seconds on the task and produced a broad report with sources.
According to the article, the result was largely accurate and gave a nuanced overview of the history. The author, who said they had written a lot about the subject over the past 20 years, considered the report stronger than most video game history books in print today.
Still, the same test exposed a familiar weakness. The report included a direct quote attributed to William Higinbotham that appeared to combine material from two sources that were not cited in the source list. The issue was not simply whether Claude found real sources; it was whether the report’s statements accurately matched the sources it presented.
That distinction matters for serious research. A generated report can be a useful starting point, but the citations, quotations, and claims still need to be checked against the original material. The source article’s conclusion was measured: Claude Research did a relatively good job on that topic, but users should dig into each source before relying on it.
Integrations bring Claude into workplace apps
Anthropic’s announcement was not limited to research depth. The company also introduced Integrations, a feature that lets Claude search connected applications in addition to web search and Google Workspace integration.
These Integrations use remote Model Context Protocol servers across web and desktop applications. Anthropic introduced the MCP standard last November as a way to connect AI applications with external tools and data sources.
At launch, Claude supports Integrations with 10 services: Atlassian’s Jira and Confluence, Zapier, Cloudflare, Intercom, Asana, Square, Sentry, PayPal, Linear, and Plaid. Anthropic also plans to add more partners like Stripe and GitLab in the future.
The practical goal is to let Claude work with information and actions inside business tools. The Zapier integration reportedly connects thousands of apps through pre-built automation sequences, which could let Claude pull sales data from HubSpot or prepare meeting briefs based on calendar entries. With Atlassian’s tools, Anthropic says Claude can help with product development, task management, and creating multiple Confluence pages and Jira work items at the same time.
Who can use the new features
Anthropic has made the advanced Research and Integrations features available in beta for Max, Team, and Enterprise users. Pro plan access is coming soon.
The company has also expanded web search, introduced in March, to all Claude users on paid plans globally. That means the broader search layer is now more widely available, while the deeper Research and connected-app features remain tied to higher-tier access for now.
The bigger takeaway is that AI assistants are moving from answering isolated prompts toward assembling documents and operating across connected systems. Claude’s upgraded Research mode may save time, and Integrations may make workplace data easier to reach. But the output still needs human review, especially when citations, direct quotes, or decisions depend on accuracy.