Google Deep Research API Updates: A Technical Breakdown

HOST

From DailyListen, I'm Alex. Google just rolled out major upgrades to its Deep Research API this week. They launched two new agents—Deep Research and Deep Research Max—through the Gemini API. Developers can now pull in open web data, mix it with their own company info, and get reports with built-in charts and infographics, all in one API call. Gemini itself powers over 750 million monthly active users worldwide, and its AI Overviews hit 2 billion users a month. But the briefing leaves gaps on pricing, exact access rules, and real-world tests of these agents. To break down what this means for developers and businesses, we're joined by Priya, our technology analyst.

PRIYA

What this unlocks is developers tapping the exact same autonomous research system Google runs inside its own products—like the Gemini app, NotebookLM, Google Search, and Google Finance. No watered-down version. You call the Deep Research or Deep Research Max agent via the Gemini API, and it fuses public web data with your private enterprise files through a single endpoint. It spits out reports with native charts and infographics, no extra coding needed. Plus, it hooks into any third-party data source using the Model Context Protocol, or MCP. Both agents crush the new DeepSearchQA benchmark at 93.3%—that's the open-source test Google dropped alongside this. Gemini's backend processes over 10 billion tokens per minute through direct API calls, so scale isn't an issue. For a dev building a market analysis tool, this means handing off complex queries and getting polished outputs fast, powered by Gemini 3 Pro's reasoning.

HOST

That 93.3% on DeepSearchQA sounds strong, but how does it stack up in practice? And those 750 million Gemini users—ChatGPT's at 810 million, Meta AI at a billion. Does this API push Google ahead?

PRIYA

The DeepSearchQA score puts both agents at the top for now—93.3% beats prior marks on that exact benchmark Google open-sourced in December 2025, when they first cracked open Deep Research programmatically. But no direct head-to-head with ChatGPT or others here; we lack those comparisons. What matters is real use. Radisson Hotel Group plugs Gemini into personalized ads and saw ad team productivity jump 50%, plus over 20% revenue growth. That's concrete. Since the Gemini 3 upgrade, AI Mode queries in Google Search run three times longer than old-school ones, showing people lean on it for depth. This API lets devs replicate that. A finance app could query stock trends, blend with internal sales data via MCP, and output charts—same as Google Finance does internally. Gemini app alone has over 650 million monthly users on its dedicated platform, up from earlier reports. But yeah, ChatGPT edges at 810 million, Meta at 1 billion. Competition's tight.

HOST

Hold on—Radisson gets 50% productivity boost and 20% revenue uptick from Gemini ads. That's real money. But for this Deep Research API, we have no customer stories yet. The briefing flags no examples of these new agents in action.

PRIYA

Right, no specific use cases for Deep Research or Max in the wild yet. That's a gap. We know the infrastructure—same as powers NotebookLM's deep dives or Search's AI Overviews, which 2 billion users hit monthly. But devs can't cite "Radisson did it" for these agents. Closest is the December 2025 pivot with Interactions API, making research programmable first time. Now with multimodal support and collaborative planning baked in, it aims to handle complex info processing better. Imagine a sales team feeding customer emails, web competitor data, and CRM files—one API call yields a report with revenue forecasts charted out. Unlocks that, but without named adopters, it's promise over proof.

HOST

Multimodal support—text, images, audio, video? How does that change reports from just text dumps?

PRIYA

Deep Research and Max now take multimodal inputs straight from Gemini 1.5 Pro's legacy setup, but upgraded with Gemini 3 Pro for elite reasoning. Feed it a video of a product demo, sales transcripts, and market PDFs—one call, and it generates a report analyzing strengths against competitors, complete with infographics plotting feature gaps. Connects via MCP to your ERP system for live sales numbers. This breaks old limits where AI choked on mixed media. Reports preserve history too—you review past runs, rerun with tweaks. Google's pitching it as your personal research assistant, available even in places like Armenia via gemini.google. For busy pros, it means less stitching sources manually. Gemini processes those 10 billion tokens per minute at scale, so no lag on big jobs. But pricing and access tiers? Undocumented here—premium enterprise likely, given Gemini 3 Pro's cost.

HOST

No pricing or access details in the briefing—that's a big hole for devs eyeing this. And Gemini 3 Pro powers it, but it's premium-priced. Any limits or risks popping up?

PRIYA

Gaps on pricing and exact access requirements mean devs test via Vertex AI or Studio tiers first, but full rollout details are missing. Gemini 3 Pro leads benchmarks for reasoning and planning—handles elite coding, long contexts up to 2 million tokens like its 1.5 Pro predecessor. But yeah, it's slower and pricier at scale than Flash variants. Risks? Same infrastructure Google trusts internally scales fine, but fusing public web with private data demands tight controls—MCP helps, yet breaches or bad blends could leak sensitive info. No controversies noted in reports, but in a market where ChatGPT holds 810 million users, over-reliance on one agent's web pulls might miss nuanced enterprise-only insights. Still, Deep Research Max pushes harder autonomy. Unveiled Monday on the Gemini API, it builds on last December's Interactions API debut.

HOST

Risks like data leaks make sense—no incidents reported, but fusing web and private files invites them. Gemini's grown fast to 750 million users this year. How'd it catch up so quick?

PRIYA

Gemini rebranded from Bard and surged in the last twelve months, hitting 650 million on the app by late 2025 per Google, now over 750 million total per TechCrunch in February. AI Overviews alone draw 2 billion monthly users—nearly double ChatGPT's 810 million. Android dominance in markets like India, Indonesia helps. What accelerated it: integrations everywhere, from Search's triple-length queries post-Gemini 3 to NotebookLM's research pods. This API extends that to devs. A marketing firm could auto-generate competitor reports, charting ad spends from public filings plus internal campaigns. Both agents support that with native visuals. No downsides in growth stats, but competition bites—Meta AI at 1 billion. Google's edge: same internal-grade agent externally.

HOST

2 billion for AI Overviews dwarfs the app's 650 million—that's wild reach. But Gemini 2.5 Pro's slower for research work. Does 3 Pro fix that?

PRIYA

Gemini 3 Pro targets research, code, and planning with top benchmark scores—like 54.6% on Humanity's Last Exam—outrunning 2.5 Pro's focus on long-context but slower speeds. 2.5 Pro suits code gen in Vertex AI, but 3 Pro handles elite reasoning for these agents. Deep Research Max likely leans on it for tougher jobs, producing those charts from multimodal mixes. Example: Upload investor deck images, web news clips, internal forecasts—MCP pulls stock APIs, outputs infographic timelines. Scales to Google's 10 billion tokens per minute. But higher costs limit to enterprise. No speed specs here, yet internal use in Finance and Search implies it's tuned for production. Gaps persist on multimodal details—no deep tech on collaborative planning.

HOST

Those token numbers show muscle, but no multimodal deep dive or planning specifics in the briefing. Let's hit that gap. What's collaborative planning mean here, even vaguely?

PRIYA

Collaborative planning lets the agent break down your query into steps, refine mid-process like a team huddle, then execute—powered by Gemini 3 Pro's planning smarts. Vague in docs, but think: You ask for "analyze EV market," it plans web scrape, competitor financials via MCP, your sales data, then charts trends. Multimodal adds video/audio analysis to that flow. No fine print on how—briefing skips it. Ties to NotebookLM's multi-doc pods, where Gemini fuses sources autonomously. You get history preservation: Rerun, tweak, review old reports. Available now in English spots like Armenia. For pros, it cuts hours off research. But without examples, it's theory. Pairs with DeepSearchQA's 93.3% validation.

HOST

Rerunning reports with history sounds handy for tracking changes. But no industry reactions—no analysts weighing in on if this beats rivals.

PRIYA

Yeah, briefing has zero expert takes or rival comparisons. AINews calls it a step forward for complex processing, VentureBeat notes the agents' web-enterprise fusion. But no quotes from devs or competitors. Gemini's wave—750 million users, 2 billion Overviews—suggests momentum, yet ChatGPT clings to top spot at 810 million. This API, from Monday's drop, gives devs Google's internal edge. Radisson's 50% productivity win shows what's possible broadly. Risks? Enterprise data in AI loops—no incidents, but watch for accuracy slips on web noise. Open-source DeepSearchQA lets anyone test rigor. Bottom line: Unlocks pro-level research tools, gaps notwithstanding.

HOST

No head-to-heads or reactions leaves us guessing on edges over rivals. Strong close though—internal power to devs. Priya, spot on as always.

HOST

One last bit: Gemini's at 750 million users, but that 1.4 billion person market with fast AI pickup—mostly mobile Android spots. Scales big, but any regional catches?

PRIYA

Strong in Android-heavy zones like Vietnam, Indonesia, India, Philippines—1.4 billion total market accelerating AI use. Gemini app's 650 million reflects that mobile-first pull. Deep Research API extends it globally via Gemini API, with early spots like Armenia live. No regional limits noted, but enterprise tiers might gate premium access. Ties back to scale: 10 billion tokens/minute handles spikes. No controversies, but in crowded field—Meta's billion, ChatGPT's 810 million—Google bets on Search integration for stickiness. Agents like Max push boundaries with charts from fused data.

HOST

Makes sense—mobile dominance fuels the numbers. Priya, thanks for laying out the unlocks and gaps.

HOST

Folks, that's the upgrade to Google's Deep Research API—powerful agents for devs, backed by Gemini's huge scale, but light on pricing, cases, and rival matchups. Check gemini.google for trials. I'm Alex. Thanks for listening to DailyListen.