ABAB News Logo
Back to Flash News

Google Set to Release Flagship AI Model Gemini 4

Google is about to release the next-generation flagship large model, Gemini 4. The head of Google DeepMind revealed this news at an AI industry summit on Wednesday. This is a key product for Google to turn the tide after falling behind Anthropic and OpenAI in the large model race.

Before the release, Google’s model schedule has repeatedly gone off track. The previous generation, Gemini 3.5 Pro, was showcased at the I/O conference in May with a promise to launch "next month," but it missed two deadlines in June and mid-July, and by August it still had not been released, totaling three delays. Internal engineers attributed the delays to a lack of prioritized investment in AI programming capabilities, with Google’s model significantly lagging behind competitors in programming intelligence benchmarks. At one point, Gemini fell out of the top 10 in usage on the commonly used OpenRouter list. Meta's Chief AI Officer Alexandr Wang publicly mocked, "Gemini who?" on X.

Following this, there was a major reshuffle in leadership. On August 6, Demis Hassabis stepped down as CEO of DeepMind to become chairman of DeepMind and chief scientist of Alphabet. Former CTO Koray Kavukcuoglu was promoted to senior vice president, responsible for the daily operations of DeepMind, reporting directly to Pichai, and took full control of the Gemini 4 development. The power center shifted from London to the Mountain View headquarters. On August 5, Jeff Dean, who had worked at Google for 27 years and was a co-leader of Gemini's technology, left along with Sanjay Ghemawat, co-leader Oriol Vinyals, and Google Brain co-founder Quoc Le to start a startup called Discovery Loop.

The talent drain has continued for months. In June, Gemini co-leader Noam Shazeer moved to OpenAI; AlphaFold co-inventor and Nobel laureate John Jumper joined Anthropic, followed by AlphaFold veterans Jonas Adler and Alexander Pritzel. Over 580 employees signed an open letter opposing Google’s defense contract with the Pentagon; in May, London employees initiated a union formation, marking the first formal attempt to form a union in a cutting-edge AI lab. Some engineers reported working 60 hours a week for a long time, with the team generally exhausted.

Google's business foundation remains large. In the second quarter, Gemini applications reached 950 million monthly active users, with daily active users doubling year-on-year; the model API processed 22 billion tokens per minute, up from 16 billion in the previous quarter; 9 million developers used the Gemini model for development each month. Google Cloud revenue was $24.8 billion, an 82% year-on-year increase; unfulfilled orders reached $514 billion, with a quarterly increase of $50 billion. The AI model for search surpassed 1 billion monthly active users. Alphabet's total quarterly revenue was $119.8 billion, a 24% year-on-year increase.

In market mechanisms, Gemini 4 bears the pressure of capital expenditure returns. Alphabet raised its 2026 capital expenditure guidance from $180 billion to $190 billion to $195 billion to $205 billion, with the second quarter marking the first negative free cash flow quarter in about 20 years. After the earnings report was released, the stock price dropped 3.65% in after-hours trading. Buyers are institutional investors and cloud customers betting on Google AI monetization, while sellers are funds worried about uncontrolled spending. Funds are flowing into TPU chips, servers, and data centers: Google has begun selling TPU computing power directly to external cloud customers and in August obtained warrants to purchase up to 58.97 million shares of Marvell stock. If Gemini 4 can recover the gap in programming capabilities, it will benefit Google Cloud and the custom chip supply chain; those under pressure are Anthropic and OpenAI, which rely on model leadership to maintain API pricing.

Supplementary data: Gemini 3 Pro was released on November 18, 2025, simultaneously launching the Gemini application, Google Search, AI Studio, and Vertex AI. Gemini 3.1 Pro was released in preview on February 19, 2026, scoring 77.1% on ARC-AGI-2, more than double that of Gemini 3 Pro; SWE-Bench Verified scored 80.6%.

Source: Public information

ABAB AI Insight

Google is the technological source of this wave of AI but has repeatedly missed opportunities to capitalize. In 2017, Google researchers published the paper "Attention Is All You Need," proposing the Transformer architecture, and all eight authors later left Google. Noam Shazeer left in 2021 to found Character.AI; in August 2024, Google brought him back with a licensing deal worth about $2.7 billion, but by June 2026, he moved to OpenAI again. In February 2023, Bard made a mistake in its first demonstration regarding a question about the Webb telescope, causing Alphabet's market value to evaporate by about $100 billion in a single day. In February 2024, Gemini had to suspend a feature due to serious deviations when generating images of historical figures. The Gemini 3 release in November 2025 temporarily put Google back in the lead, triggering an internal "red alert" at OpenAI, but the lead lasted less than six months.

On the capital path, Google has made three layers of bets. The first layer is building a full stack: from TPU chips, Google Cloud to the Gemini model, and then to distribution channels like Android, Chrome, and Search, with annual capital expenditures of about $200 billion. The second layer is hedging: Google holds about 14% of Anthropic's shares, with cumulative investments of about $3 billion; in October 2025, Anthropic signed a power agreement to use up to 1 million TPUs, worth billions of dollars. The result is that Google is supplying computing power and funds to competitors while watching its core researchers being poached. The third layer is selling TPUs to external customers, turning chips from a cost center into a profit center, so even if it loses in the model race, it can still collect rent from competitors' training needs.

The closest analogy is Microsoft's missed opportunity in the mobile internet: in 2013, Microsoft acquired Nokia's mobile business for $7.2 billion, and in 2015, it wrote down about $7.6 billion. It wasn't until Nadella shifted focus to cloud computing that Microsoft turned around. Intel also saw its chip manufacturing process surpassed by TSMC, leading to CEO Gelsinger's forced departure in December 2024. Google's attempt to counter Facebook with Google+ in 2011 is another example of a giant holding channels but failing to catch up in a new race. Google is currently in a phase of "maintaining channels while chasing models": 950 million monthly active users prove its distribution capability, while the decline in OpenRouter rankings indicates developers are voting with their feet.

The essence is the transfer of pricing power. In the AI industry, pricing power is shifting from those who control distribution channels to those who possess the strongest model capabilities, with programming intelligence being the most typical battleground. The mechanism is that programming intelligence directly replaces the labor costs of engineers, and a company's willingness to pay is closely tied to the model's capabilities; a few percentage points of capability difference can lead to significantly different outputs. The switching cost for APIs is extremely low, allowing developers to switch suppliers overnight, resulting in revenue concentrating heavily on the top-ranked models. Even with search and Android, Google cannot lock users into a second-rate model. Gemini 4 must regain its position in programming capabilities; otherwise, $200 billion in capital expenditures will only yield computing power rent, not pricing power.

ABAB News · Cognitive Laws

  1. Channels can buy users, but cannot buy leadership.

  2. The inventors of technology may not be the ones to reap its benefits.

  3. In the AI race, delaying a quarter is equivalent to giving up a generation.