While OpenAI and Anthropic compete to create recursively self-improving systems capable of independently designing the next generations of algorithms, Google DeepMind has chosen a fundamentally different path. Externally, the difference may seem insignificant — the corporation continues to release models and make money. However, an analysis of Demis Hassabis's division's recent steps points to a strategic shift towards so-called "world models."

Why Gemini 3.6 Flash is Not About Power

On July 21, Google released the Gemini 3.6 Flash model. The key highlights are speed and cost, not absolute performance. According to the corporation itself, the new version generates 17% fewer tokens than its predecessor, which directly reduces the cost per response. However, in the Artificial Analysis ranking, this model only placed 10th, losing to virtually all leading competitors.

This does not mean Google has abandoned refinements. The company has already launched the largest training run for the future Gemini 4, and a larger-scale version of Gemini 3.5 Pro is currently being tested with partners. But the main point is that Sundar Pichai publicly linked the roadmap to the development of personal agents, not to positions in benchmarks.

DeepMind vs. Recursive Improvement

A world model is an AI that learns the physical patterns of an environment: gravity, motion, cause-and-effect relationships. Such an algorithm predicts the unfolding of events in space, rather than just selecting words. On the DeepMind website, both Genie 3 and Gemini Robotics are located in the "world models" and embodied AI section — software that controls machines.

OpenAI and Anthropic have chosen a completely different strategy. These teams are focused on recursive self-improvement of systems. Essentially, this technology allows AI to independently design the next generation of algorithms, launching a development cycle where the product creates even more powerful solutions.

Anthropic co-founder Jack Clark, in his essay dated May 4, estimated the probability that AI could independently manage the research process by the end of 2028 at 60%. At the same time, he noted that DeepMind "looks the most restrained among the big three." Restraint here implies caution. Clark bases this on DeepMind's own documents, particularly a paper on AI safety issues published in 2025.

For comparison: Anthropic has already recorded cases where AI creates a more advanced AI. During one testing period in April, the performance of their models increased 52-fold. A year earlier, this increase was only 2.9-fold. An experienced engineer would need four to eight hours of work for a fourfold increase.

Why You Can't Say Google Has Left the Race

There are two important factors contradicting this hypothesis. First, Google holds the lead in machine learning research tests. The MLE-Bench benchmark evaluates models' ability to independently create AI architectures. In February, Gemini 3 showed a result of 64.4%, taking first place. Later, in July, version 3.6 Flash scored 63.9%. Companies exiting the market usually do not top such rankings.

Second, the IT giant maintains enormous reach. Sundar Pichai stated that the audience for the Gemini app reached 950 million people per month.

The Financial Cost of a Waiting Position

Search advertising revenues have long covered DeepMind's experimental expenses. However, recent Alphabet reports indicate a shrinking financial cushion. In the second quarter of 2024, the company's revenue reached $119.8 billion, 24% higher than a year earlier. Search advertising alone brought in $63.3 billion.

But costs are rising. Over three months, Alphabet allocated $44.9 billion for data center development and equipment purchases — almost double the amount from a year ago. As a result, free cash flow was negative $5.86 billion. In March, this figure was positive at $10.1 billion, and in December of the previous year, it was $24.6 billion.

To cover the shortfall, Alphabet issued $49.6 billion in shares and raised $20.3 billion in loans. Long-term liabilities grew from $46.5 billion to $98.2 billion over six months. The expense item related to joint AI developments resulted in a loss of $5.79 billion — a year earlier it was $3.37 billion. Consequently, the current waiting position is costing the corporation increasingly more.

What to Watch in the Next 30 Days

  • Whether the release of Gemini 3.5 Pro will take place and what positions the model achieves in tests.
  • Whether DeepMind will present new data on world models for Gemini 4.
  • Whether Alphabet's cash flow will return to positive values in September.
  • Whether Demis Hassabis will articulate a clear position regarding self-learning technology.

The launch of Gemini 4 will be a defining moment for the company's strategy. If the approach to creating world models proves itself, the current slowdown will turn out to be a precise calculation. The upcoming financial report will show the real limits of Alphabet's resources in this technological race.

Expert opinion: While the market is fixated on benchmarks and recursive improvement, DeepMind is making a long-term bet on understanding the physics of the real world. If this strategy pays off, Google could find itself in a unique position where competitors are catching up not on tokens, but on fundamental understanding of the environment. However, the cost of error in the form of growing debts and departing talent is becoming critical.