Google has pulled its Gemini Deep Research agent out of the consumer app and repackaged it as a developer-ready product. On Wednesday, the company announced two new versions — Deep Research and Deep Research Max — now in public preview through the paid tiers of the Gemini API. The big deal: for the first time, the agents can combine open-web searching with a company’s private data streams in a single API call, then return a fully cited, chart-embedded report.
In other words, Google is no longer selling you a smarter chatbot. It is selling the research analyst itself.
What actually changed
The original Deep Research agent, released to developers in December 2025 via the Interactions API, was essentially a sophisticated summariser. The new versions are built on Gemini 3.1 Pro, Google’s most factual model to date, and have been retrained to plan multi-step investigations, navigate paywalled databases, and stitch the findings together with inline visualisations.
Google is splitting the product into two flavours, because research workloads aren’t all the same shape:
- Deep Research is the faster, cheaper option. It’s tuned for interactive surfaces, like a chat window where a user expects a report within seconds rather than minutes. It replaces the December preview.
- Deep Research Max is the opposite. It uses extended test-time compute to keep reasoning, searching, and refining until it has produced an exhaustive report. Google’s pitch is a nightly cron job that drops a full due-diligence brief on an analyst’s desk by morning.
For heavy jobs, the resource footprint is considerable. Google’s own developer documentation estimates a complex competitive-landscape task can burn roughly 160 search queries, 900,000 input tokens, and 80,000 output tokens per run. That is not a query; that is a small consulting engagement rendered as an API call.
The private-data piece is the real story
The headline capability most enterprises will care about is the ability to point the agent at gated, proprietary information. Deep Research now supports the Model Context Protocol (MCP), which means developers can plug it into their own internal data sources, file stores, or subscription data feeds. Google says it is working directly with FactSet, S&P, and PitchBook on MCP server designs so paying customers can blend those financial databases into agent workflows.
You can also flip the behaviour entirely: turn web access off and restrict the agent to internal files only. That matters for regulated industries like banking, insurance, and healthcare, where letting an AI roam the open web while handling sensitive data is a non-starter.
Other additions include collaborative planning (you review and tweak the agent’s research plan before it runs), multimodal input (PDFs, CSVs, images, audio, video), inline charts and infographics inside the final report, and live streaming of the agent’s intermediate thinking.
This is a direct continuation of the agent-first strategy Google signalled when Gemini 3 launched in November, and a clear shot at OpenAI’s ChatGPT Deep Research and Anthropic’s Claude research mode, both of which now support similar MCP-style private-data hookups.
Read the benchmarks with one eye closed
Google claims state-of-the-art numbers on three retrieval-and-reasoning benchmarks: Humanity’s Last Exam, DeepSearchQA (Google’s own open-sourced test), and BrowseComp. But as The Decoder pointed out, competitors report higher numbers on the same tests when they run them through their own tooling. OpenAI says GPT-5.4 Pro hits up to 89.3% on BrowseComp. Anthropic claims 84% for Opus 4.6, achieved with reasoning turned off. Google’s figures aren’t fabricated, but methodology differences between “raw API” and “wrapped in our own scaffolding” make cross-lab comparisons slippery. Take the leaderboard with salt.
What this means for Kenyan and African teams
For developers and small analytics shops locally, the immediate implication is that serious research automation no longer requires a data-science team. A Nairobi-based fintech could, in theory, wire its internal customer database into Deep Research Max and get overnight competitor briefs that previously demanded a junior analyst’s week. The same goes for law firms doing due diligence, media houses fact-checking long investigations, or university research departments short on staff.
The cost question is still open. Google has not yet published finalised pricing, only that the agents run on paid tiers. Given the token volumes involved, a single complex run could land in the tens of US dollars — not trivial when converted to KES at scale, but still dramatically cheaper than billable consultant hours.
The safety asterisk
Google is upfront that pointing an agent at both the open web and private files amplifies attack surface. Its own documentation warns about prompt-injection attacks hidden inside uploaded files and malicious web pages engineered to manipulate the agent’s output. The advice is standard but worth repeating: trust your file sources, and always verify citations in the final report.
Deep Research and Deep Research Max will roll out to Google Cloud’s enterprise customers “soon.” Until then, they live inside the Gemini API’s Interactions endpoint, available to any developer willing to pay.
