Gemini 3 Pro launched Nov. 18, 2025 as a sparse mixture-of-experts reasoning model handling up to 1 million input tokens natively, and Google spent the following six months converting that architecture into shipped product 1. By Google I/O 2026, held May 19 and 20, Google had rebuilt Search itself around "Search Agents and Generative UI," a description the event's own materials used for what amounted to the largest interface overhaul the product had received in years 3. Momentum this visible carries a cost few outsiders see: three days after Alphabet's second-quarter earnings closed the books on Aug. 5, 2026, Jeff Dean and three senior colleagues walked out of Google to found a rival science venture, taking decades of institutional research history with them.
The Reasoning Race Accelerates
Gemini 2.5 Pro Experimental set the stage in March 2025, topping the LMArena leaderboard and posting strong results on Humanity's Last Exam while pairing chain-of-thought prompting with native multimodality 1. Flash and Flash-Lite variants followed that spring, and Google I/O 2025 named 2.5 Flash the default model across its consumer surfaces 1. Gemini 3 Pro's November arrival escalated the architecture itself: a sparse mixture-of-experts design lets the model activate only a fraction of its parameters per query, a structural choice that keeps inference costs manageable even as context windows stretch to a million tokens 1. December completed the family with Gemini 3 Deep Think and Gemini 3 Flash, giving Google a three-tier lineup spanning maximum reasoning depth down to low-latency response speed within a single month 1.
Nano Banana, Gemini's image-generation feature, offered an early proof point that model quality translates into adoption: the tool pulled more than 10 million new users into the Gemini app within weeks of its August 2025 launch 1. That kind of viral single-feature growth gave Google evidence its research advances could move consumer behavior directly, a pattern the company leaned into again at I/O 2026.
Silicon underlies every part of that cadence. Google's Ironwood TPU generation, its seventh, entered production specifically tuned for inference at scale rather than training alone, a design choice that matters more once a company ships reasoning models handling million-token contexts across hundreds of millions of daily queries. Custom silicon of that kind gives Google a cost structure distinct from what rivals renting third-party chips can match, a structural advantage that compounds every time Gemini's release cadence tightens.
Search Learns to Act
Google I/O 2026 pushed Gemini's reasoning capability into the product Google depends on most. "Search Agents and Generative UI" replaced static results with interfaces that assemble themselves around a query and can complete multi-step tasks on a user's behalf, a shift the event framed as central rather than experimental 3. Alongside the Search overhaul, Google introduced Gemini 3.5 Flash, Gemini Omni — a multimodal model capable of generating video output from any input type — and a companion feature called Gemini Spark 23. A Universal Commerce Protocol for agent-driven shopping extended the same agentic logic into transactions, letting a Gemini-powered agent complete a purchase rather than merely recommend one 3. Google also announced AI-powered audio glasses built with Samsung, Warby Parker and Gentle Monster, slated for a fall 2026 launch, extending Gemini's reach past the browser and into a wearable form factor 3.
Read together, the announcements describe Google treating agentic capability as a product requirement across every surface it controls, distinct from a single flagship chatbot competing on benchmark scores. Search Agents in particular puts Gemini's reasoning directly inside the revenue engine Google has protected since the 1990s, a bet that agentic search sustains advertising and commerce revenue better than a static results page facing pressure from ChatGPT and Perplexity.
Dean's Exit and the Discovery Loop Bet
Jeff Dean's departure landed with weight proportional to his tenure: Google's chief scientist and 30th employee, a researcher whose fingerprints touch Search infrastructure and Gemini's own development, left the company he helped build 4. He carried three senior colleagues with him — Sanjay Ghemawat, a senior fellow and one of Google's most decorated engineers; Quoc Le, a founding member of Google Brain; and Oriol Vinyals, a senior research scientist at Google DeepMind 4. Their new venture, Discovery Loop, structured itself as a public benefit corporation aimed at automating scientific experimentation, running what the founding team described as a higher quantity and quality of experiments than human-paced research allows 4.
Funding for Discovery Loop came from Radical Ventures and Khosla Ventures as co-leads, with Kleiner Perkins, Lightspeed and Doerr Capital participating — and notably, Alphabet itself provided financial backing to the venture its own researchers just left 4. That detail complicates any reading of the departure as adversarial. Alphabet investing in the company its former chief scientist just founded suggests a negotiated exit rather than a rupture, closer to a spinout Google chose to fund than a defection it failed to prevent. Demis Hassabis remains atop Google DeepMind's research organization as chief executive, steering the Gemini roadmap through exactly the period Dean's team chose to leave, and the timing — three days past earnings, four researchers at once — reads as a coordinated transition rather than a scattered exodus.
Reading the Two Curves Together
Google's product momentum and its research-talent churn moved on the same calendar, pulling toward different destinations. Gemini 3 Pro shipped in November, I/O 2026's Search Agents shipped in May, and Discovery Loop launched in August, each event roughly ninety days apart, forming a rhythm distinct from the multi-year gaps that once separated major Google AI announcements. Dean's team departing amid that rhythm, rather than before or well after it, suggests Google's research organization reached a natural inflection point where the infrastructure generation Dean represented handed off to whatever comes next under Hassabis.
Alphabet's willingness to fund Discovery Loop rather than fight its founding hints at a broader strategic calculation: keeping a research relationship alive through investment costs less than losing access to Dean's future work entirely, and Google potentially gains a scientific-discovery partner it might otherwise have spent years building in-house. Whether that calculation pays off depends on Discovery Loop delivering results distinct enough from Google's own research output to justify the arrangement, a question only time and published findings can answer.
By the numbers
- 1 million: maximum input tokens Gemini 3 Pro handles natively, per its Nov. 18, 2025 launch specifications 1.
- Three: Gemini 3 variants (Pro, Deep Think, Flash) Google shipped across a single month, November into December 2025 1.
- 10 million: new Gemini app users Nano Banana attracted within weeks of its August 2025 launch 1.
- May 19-20, 2026: the dates of Google I/O 2026, where Search Agents and Generative UI debuted 3.
- Four: senior Google researchers, led by Jeff Dean, who departed Aug. 5, 2026 to found Discovery Loop 4.
- 30th: Jeff Dean's employee number at Google, a tenure spanning more than two decades before his exit 4.
What to watch
Discovery Loop's first published results will test whether Dean's team can prove automated scientific discovery works at the scale their pitch promised, a benchmark distinct from any language-model leaderboard. Gemini's Search Agents rollout, tracked against user complaints or adoption metrics through the rest of 2026, will show whether Google converted its architectural lead into a product enterprise buyers and consumers trust with transactions. Hassabis's next public roadmap statement, whenever DeepMind issues one, would clarify how the research organization plans to backfill the specific expertise Dean, Ghemawat, Le and Vinyals carried out the door.
Sources
- Wikipedia contributors, "Gemini (language model)," Wikipedia, accessed Sept. 4, 2026, https://en.wikipedia.org/wiki/Gemini_(language_model)
- Wikipedia contributors, "Gemini (chatbot)," Wikipedia, accessed Sept. 4, 2026, https://en.wikipedia.org/wiki/Gemini_(chatbot)
- Wikipedia contributors, "Google I/O," Wikipedia, accessed Sept. 4, 2026, https://en.wikipedia.org/wiki/Google_I/O
- Lucas Ropek, "Jeff Dean and other top AI researchers are leaving Google to launch their own startup," TechCrunch, Aug. 5, 2026, https://techcrunch.com/2026/08/05/jeff-dean-and-other-top-ai-researchers-are-leaving-google-to-launch-their-own-startup/