● Google Unleashes Gemini Deep Research API, Search Dies, Agent Research Takes Over
Google Opens the ‘Gemini Deep Research’ API: The Market Is Shifting from “Search” to “Automated Investigation (Research Automation)”
In this post, it’s not just “Google improved an AI research tool” level,
① how the architecture of autonomous research (Deep Research) that digs deep into the web has changed,
② why the ‘Interactions API,’ built so developers can embed it directly into apps, is a game changer,
③ why the newly released benchmark DeepSearchQA becomes the battleground of the “AI agent era”,
④ immediately monetizable application scenarios in high-precision industries such as finance and bio,
⑤ (important) Google’s long-term strategy and risk points that other news rarely mentions
I’ll organize all of this in one go.
1) News Briefing: Today’s Announcement, Key Points Only
[What happened?]
Google completely redesigned ‘Gemini Deep Research’ and opened it to developers via an API.
The reasoning engine uses the latest model, Gemini 3 Pro.
[Why does it matter?]
This is not “AI for writing reports,” but an event that enables you to embed a ‘research agent’ into your service that maintains long sessions and repeatedly executes plan → search → verify → re-search.
In other words, information discovery centered on a search box is moving to agent-centered automated investigation.
[Core update points]
1) Web search performance greatly improved: explores “deep areas” inside sites to find specific information
2) Released the ‘Interactions API’: supports long sessions/server state/long-horizon reasoning loops running in the background
3) Released the open-source benchmark ‘DeepSearchQA’: focuses on evaluating realistic multi-step web research performance
2) Functional Change: Leveling Up from a “Research Tool” to “Research Infrastructure”
2-1. What Deep Research does (how it works)
Gemini Deep Research is not a chatbot that simply takes a question and answers it.
It creates its own research plan,
then repeats search query generation → reading results → identifying knowledge gaps → re-search.
The key point here is that it’s not “search once and done,” but is designed around long-horizon reasoning.
2-2. Why “minimizing hallucinations” is the key takeaway
Google said it specially trained this agent to minimize hallucinations and maximize report quality.
From a company’s perspective, the scariest thing in research automation is a “plausible but false report,” and this direction aims to reduce that risk head-on.
2-3. What it means that web search got “deeper”
The improvement mentioned in the article is not just better search accuracy, but the ability to explore deep inside a website to find specific data.
This is important.
Most practical information (full-text regulatory documents, appendix materials in corporate IR, research appendix data, detailed public-data pages) is not on the “surface page.”
3) Interactions API: An “Agent Operating System” Opened to Developers
3-1. How is it different from existing APIs?
Typical AI APIs often have a “request → response” structure that ends in one shot.
This ‘Interactions API’ is the opposite.
It supports maintaining long sessions, managing server state, building multi-step plans, and running long-horizon reasoning loops in the background.
3-2. What becomes genuinely easier in real work
Developers can provide PDFs, CSVs, documents, web links, datasets, and more all at once,
and the agent judges for itself “what’s missing” and fills gaps through re-search.
If this runs properly, a company’s research pipeline changes from “human hands” to an “agent workflow.”
3-3. Meaning from an economic/industry perspective
At this point, digital transformation steps up a level.
Now corporate competitiveness shifts from “how fast you can find information” to
how reliably you can operate and control AI agents (governance/security/auditability).
4) Real Use Cases: Finance Due Diligence and Bio Make Money First
4-1. Finance: Automating due diligence and market analysis
According to Google, it is already being used in high-precision fields like financial due diligence and market analysis.
Financial firms automatically collect and organize
market signals, competitor information, and regulatory risks across the web plus internal materials.
The practical point is this.
Research cost ultimately equals “people’s time,” and
if an agent handles first-pass data collection plus summarization/organization,
analysts can focus on “interpretation and decision-making.”
This significantly impacts productivity from a corporate cost-structure perspective.
4-2. Bio: Biomedical literature analysis → accelerating early-stage drug research
In bio, massive volumes of papers/reviews/meta-analyses/adverse-event data are core, and
Deep Research reportedly analyzes literature at scale to pull forward early-stage drug development research.
One reason bio is getting attention again in today’s market is “can AI shorten R&D lead time,” and this aligns exactly with that direction.
5) DeepSearchQA Benchmark Released: Moving from “Getting the Answer” to “Reproducible Investigation”
5-1. Why create a new benchmark?
Google released the open-source benchmark DeepSearchQA, and said it focuses on evaluating the complexity of realistic multi-step web research that existing tests failed to capture.
5-2. Composition: 17 domains, 900 ‘causal chain’ tasks
It’s not a simple quiz; it’s structured so that
cause → effect → reasoning progresses across multiple steps, making it good for assessing “investigation capability.”
5-3. The measurement method changed
It measures not only correctness but also answer comprehensiveness and search recall.
This is a huge change from a corporate perspective.
In work settings, “one correct line” matters less than “how completely you gathered the evidence.”
5-4. Results: Deep Research 66.1%
On DeepSearchQA, Gemini Deep Research recorded 66.1%,
about 10%p higher than Gemini 3 Pro (56.6%),
and the article says it surpassed OpenAI GPT-5 Pro (65.2%).
5-5. Also top-tier on other benchmarks
HLE (Humanity’s Last Exam) 46.4%
BrowseComp 59.2%
It says these were the highest scores among the compared models.
6) Competitive Landscape: “#1 on a Benchmark” Lasts a Day, but the Game Has Changed
Right after Google’s announcement, OpenAI released GPT-5.2 and there were reactions that some benchmark rankings flipped.
It’s also true that the industry cynically says “benchmark superiority is only valid on announcement day.”
But the essence of this issue is not a score race; it’s this.
The war over APIs/platforms to deploy and operate agents has entered the main game.
That is, it’s shifting from who is smarter to who can be embedded more deeply “inside workflow.”
This trend also matters in AI investment themes.
Rather than only looking at model performance, you need to look at the platform lock-in created by API ecosystems/developer tools/integrated product suites.
This ultimately affects big tech’s cashflow-driven CapEx and the reshaping of enterprise IT spending, independent of interest rates.
7) Google’s Next Move: Integration with Search, Finance, NotebookLM, and the Gemini App Is the Truly Scary Part
Google said it plans to integrate Gemini Deep Research
into Google Search, Google Finance, NotebookLM, and the Gemini app in sequence.
What this integration means is simple.
It changes UX from us typing queries and opening links,
to AI exploring and analyzing on our behalf, then delivering it as a “conclusions + evidence package”.
Especially when combined with Google Finance, market participants can have a flow where
macro indicators, earnings, and regulatory issues are connected and automatically researched all the way to “what does this mean for my portfolio.”
This is effectively an event that dramatically lowers the “access barrier to research” even for retail investors and office workers.
8) The “Most Important Point” Other News Rarely Mentions (Core Blog Takeaway)
8-1. Google is targeting not “search share,” but the “standard for verification”
If an era comes where people search less, search traffic itself could decline.
But rather than fearing that, Google seems strongly focused on
making the “verification loop” and “evidence exploration” that AI must pass through during research into a standard on its platform.
The combination of Deep Research + DeepSearchQA (publishing evaluation criteria) is also a way of setting the rules for
“this is how agent research should be evaluated.”
8-2. Interactions API opens the market for “AI agent backend”
Most people look only at model performance, but real work is different.
Without long-running jobs, state management, failure recovery, retries, and log/audit tracking, enterprise adoption gets blocked.
This API is a declaration that the platform will take on that “messy real-world work” at the platform level, so the impact is large.
8-3. Rather than benchmark numbers, optimizing “thinking time” becomes the revenue model
Google said it can analyze whether “more thinking time” leads to real performance gains.
This connects directly to the pricing model.
The longer an agent thinks, the higher the cost, and
enterprises must find the optimum point of “accuracy vs cost.”
Going forward, AI adoption budgets may shift away from software subscriptions and toward a larger share of variable costs based on
research workload (inference time/search calls/verification loops).
8-4. Risk: The more “deep web exploration” increases, the more data/copyright/regulatory issues grow
The ability to go deep into a website and collect data
also means, from a service operations perspective,
issues around copyright/robots exclusion policies/data usage rights/personal information/regulatory compliance may erupt more frequently.
When a company adopts Deep Research, before technology, it must ensure
data governance (how far collection is allowed), audit logs, and source management move together.
9) Practical Implementation Guide: How to Attach It to “Our Company/My Work”
9-1. Top 5 tasks with immediate impact
1) Competitor/industry trend monitoring: automatic weekly research reports
2) Regulation/compliance checks: detect policy changes + summarize impact
3) Investment/business feasibility reviews: organize market size, customer segments, risk chains
4) Technology scouting: mapping paper/patent/open-source trend landscapes
5) Research combining internal documents + external materials: connect internal knowledge and market evidence into one document
9-2. Adoption checklist
– Can you enforce source citations/evidence links in the output?
– Can you control the maximum number of re-search/re-verification loops (cost)?
– When mixing internal documents (PDF/CSV) and the external web, are permissions/security separated?
– Can you institutionalize a policy that the output must state “uncertainty/gaps”?
10) One-Line Macro Conclusion: AI agents lower the “cost of information” and increase the speed of industrial decision-making
This update is less “AI got smarter” and more an event that
lowers the unit cost of information discovery/verification/report generation, thereby accelerating decision-making speed.
As this trend grows, corporate field operations will undergo redesign across
research/planning/strategy/IR/legal/RA, and
in markets, platform lock-in around AI infrastructure and platforms will strengthen, making the big-tech-centered AI investment narrative more solid as well.
< Summary >
Google overhauled ‘Gemini Deep Research’ and opened it via API, fully ushering in an era of “research automation” where AI plans for itself and deeply explores and verifies the web.
The key takeaway is the Interactions API that supports long sessions and background inference loops, and it delivers immediate value in high-precision work such as financial due diligence, market analysis, and bio literature analysis.
Releasing the DeepSearchQA benchmark is an attempt to make “reproducible investigation” the standard over mere “correct answers,” and the nature of competition will move from model scores to agent operations platforms and integrated ecosystems.
[Related posts…]
- The spread of Gemini-based AI agents and how workflow automation reshapes corporate competitiveness
- The AI agent era: A summary of trends reshaping search, research, and decision-making
*Source: https://www.aitimes.com/news/articleView.html?idxno=204767
