Gemini 4 Argon delivered a debut report card that even Google was unprepared for: within 72 hours of launch, it processed nearly 25 trillion tokens, with consumption reaching 3 times that of the previous generation Gemini 3.0 over the same period. Even Google, with its deep pockets, had to activate compute throttling once again in the face of the flood of incoming requests. In a recent interview on the tech podcast Sources, Google CEO Sundar Pichai walked through the details of this release and, for the first time, laid out clearly the things that had most puzzled outsiders over the past few months — what happened to Gemini 3.5 Pro, which had been name-dropped at the I/O conference; why DeepMind's leadership suddenly reshuffled; and what the next-generation model actually looks like.
The answer to 3.5 Pro's disappearance was surprisingly blunt: the team voluntarily abandoned it to seize the high ground in Agent coding capabilities. In the middle of this year, while training the 3.5 series, Google discovered that the model had an obvious gap in Agentic Coding. The team judged that rather than forcing out a larger Pro, iterating rapidly with a smaller, more efficient Flash version would be the fastest climbing path to close the shortcoming. So over roughly three months, they ran an intensive series of consecutive leaps from 3.5, 3.6, 3.7, to 3.8 Flash. By the time this approach was validated and matured, ready to be ported to Pro, the pretraining of Gemini 4.0 Argon had already quietly been completed.
In a compute-constrained environment, releasing another 3.5 Pro with an extremely short lifecycle clearly wasn't worthwhile, so the team injected all the technical know-how validated through Flash directly into Argon's post-training stage. This decisive cut brought Argon a step change in long-horizon tasks — in external testing, it autonomously beat Minecraft after a little over a day and less than 2000 dollars of compute; within Google, it has already been pointed at the Chromium codebase with millions of lines of code. As for the rumored Carbon model, Pichai clarified that it was merely a codename for internal frontier iterations, not a standalone new product; the team is currently distilling based on Argon, developing a lighter-weight version with similar capabilities and lower cost, to be released to more users.
Even more intriguing is the change in the R&D approach itself. Pichai announced that Google internally has already entered the recursive self-improvement (RSI) stage: models are no longer just the objects of training, but directly participate in proposing training recipe suggestions and automatically evaluating the next checkpoint. A healthy collaborative relationship has formed between researchers and models, and the whole continuum is quietly shifting. He himself has become a beneficiary of this new process — every day he runs Agents to summarize the 24-hour progress of pretraining, mid-training, post-training, and the data groups, saving about an hour of information syncing time, and uses Agents connected to telemetry data to monitor each department's compute efficiency in real time. Pichai judges that 2027 will be the golden age of productivity, and within the next 12 months, the penetration of RSI will bring sustained acceleration.
The tremors at DeepMind also received their first official footnote in the interview. Jeff Dean, the legendary engineer who spent a full 27 years at Google, left to found Discovery Loop and embark on new explorations of a new stage of his life; Demis Hassabis stepped back to Chief Scientist and Chairman, focusing his energy on AI drug discovery, life sciences, and robotics deployment represented by Isomorphic Labs; taking over the model R&D pipeline is Koray Kavukcuoglu, who has been at DeepMind for more than a decade and moved to the Bay Area 18 months ago to push cross-department integration. Pichai's core explanation was straightforward: frontier model R&D has evolved into a hyper-scale engineering effort that consumes one's full attention — from data preparation and compute scheduling to algorithm evolution and coordinating collaboration between human teams and model Agents, the pace is so fast that it requires a dedicated leader putting in full effort. The Gemini app was born inside DeepMind in the first place, and deep coupling between the two sides has long been the norm; this reshuffle is more like formally marking a milestone in a merger that had already happened.
The confidence behind the aggressive investment is written on the flip side of the financial ledger. This summer, Google recorded its first negative free cash flow since its IPO, while continuing to raise its 2027 capital expenditure guidance and has already begun planning land and power reserves for 2028 to 2029. Pichai did the time math: the lead times for in-house chips and data centers are extremely long, and the compute levels the market is assessing today with perfect information are actually the result of decisions made two years ago — and at this time last year, the explosion of Agentic Coding hadn't even happened yet. He emphasized that today's AI infrastructure is by no means the idle dark fiber bubble of yesteryear; all compute is being genuinely consumed — demand signals from consumer and business sides, cloud business long-term contract backlogs, and underlying R&D breakthroughs together constitute his certainty. He even teased in the interview: based on internal benchmarks and evolution curves, the next-generation Gemini 5.0 will be far stronger than the current 4.0 Argon.
On product architecture, Google is erasing the boundary between enterprise and consumer Agents. Pichai revealed that whether it's the B-side or C-side Gemini Agent, what runs underneath is the same model, the same runtime framework, and the same runtime — the only difference is whether enterprise context or personal context is injected. The enterprise design of automatically assigning an independent micro-VM to each task to achieve 7×24-hour secure operation is being rolled out to the consumer side in sync; Gemini's existing "Chat" and "Tasks" will also be merged into a single input box. As for whether Search and Gemini will merge, his answer is: there's no need to forcibly fuse the two products — letting an Agent layer running in the cloud, online 24 hours a day, naturally share user context will allow seamless continuation of tasks across Search, the Gemini app, and Gmail. At the underlying level of web interaction, Browser Use, which operates interfaces like a human, will coexist long-term with native protocols like MCP and UCP — on one side an elegant universal standard, on the other a progressive path where users always retain a sense of control.
The boldest step is still in space. The first pilot of "Project Suncatcher," a collaboration between Google, Planet Labs, and SpaceX, has already launched; Pichai said the chips should still be running normally in orbit, and he will soon be able to say hello to that batch of space TPUs. From first principles, sun-synchronous orbit can provide 24/7 uninterrupted energy, which is a way out beyond the hard constraint that the Earth's land and power will eventually hit. He compared this to the launch of Waymo back in the day — the people who flew to the desert to watch the DARPA autonomous driving challenge back then could hardly have imagined what autonomous driving looks like today, and 10 to 20 years is enough for early steps to grow into astonishing milestones. He also placed quantum computing within a tangible future: over the past nearly decade, the team has steadily achieved or even exceeded internal milestones, and he genuinely sees it demonstrating exciting practical commercial applications within 5 years — just as Google Cloud offers TPU compute today, five years from now Google Cloud will sit down to close deals on quantum deployments.
Looking at the nearer-term picture, Waymo has just crossed the threshold of 500,000 paid rides per week, and the focus for the next 12 months is expanding to international markets, more weather conditions, and full highway scenarios while maintaining a safety standard of 95% fewer severe accidents than conventional driving, with 10 new cities launching this year; Wing drone delivery plans to greatly expand scale and broaden its categories in 2027, by which time about 40 million American households will be able to use it. On regulation, Pichai advocates establishing unified national standards rather than scattered state-by-state rules, and believes the industry should establish clear guidelines and independent audits the way the financial world does — even if what exists now is merely voluntary testing submissions to the US government, the direction is already completely different from a year ago. And on the most critical compute allocation decisions, he drew a red line: all decisions strictly involve humans (Human-in-the-loop), with models only providing input — Google was already using AI to optimize data center energy consumption five years ago, but the day that life-and-death authority is handed to models, humans will have long since moved on to the next layer of problems.