While OpenAI and Anthropic compete to create recursively self-improving intelligence, Google DeepMind has chosen a fundamentally different path. The corporation has stopped chasing leadership in large language model benchmarks and focused on "world models"—systems that understand the laws of physics. This is a strategic risk that could either redefine the industry or leave Google behind.

Externally, Alphabet continues to release products: on July 21, the Gemini 3.6 Flash model was released. Its key metrics are speed and generation cost. The token share has been reduced by 17%, lowering output costs for users. However, in the Artificial Analysis ranking, this model only placed 10th. All leading competitors outperformed it.

DeepMind's New Philosophy

Google CEO Sundar Pichai has publicly shifted priorities: the focus is now not on leaderboard positions, but on developing personal agents and "embodied AI." DeepMind has already integrated Project Genie with Street View and introduced the SIMA 2 agent, which learns by interacting with 3D worlds.

The essence of the approach is that the model should predict the development of events in space, rather than simply selecting the next word. This fundamentally differs from the strategy of OpenAI and Anthropic, which bet on recursive self-improvement (RLHF and automatic code generation).

Recursive Self-Improvement: Competitors' Trump Card

Anthropic co-founder Jack Clark estimates a 60% probability that AI will take over research management by the end of 2028. His lab is already recording cases where Claude writes over 80% of deployed code. In April, their model performance increased 52-fold in a month. For comparison, an experienced engineer requires 4–8 hours for a fourfold increase.

DeepMind, according to Clark, "looks the most restrained among the big three." Caution, backed by publications on AI safety, could result in losing momentum.

Financial Pressure

Alphabet maintains a colossal cash flow: in the second quarter of 2024, revenue reached $119.8 billion (+24% year-over-year). However, infrastructure expenses soared to $44.9 billion per quarter—almost double the previous year. Free cash flow turned negative at -$5.86 billion, and long-term liabilities grew from $46.5 billion to $98.2 billion.

The expense line for joint AI developments resulted in a loss of $5.79 billion, compared to $3.37 billion a year earlier. The wait-and-see position is costing the corporation increasingly more.

What to Watch in the Next 30 Days

  • The release of Gemini 3.5 Pro and its positions in independent tests.
  • New DeepMind data on world models for Gemini 4.
  • Whether Alphabet's free cash flow returns to positive values in September.
  • A clear position from Demis Hassabis regarding self-learning technology.

Expert Opinion: Google has made a deliberate bet on long-term understanding of the physical world, rather than short-term dominance in chatbots. If DeepMind's world models live up to expectations, the current lag in rankings will prove to be a tactical maneuver. However, financial reports show that even a giant's resources are not unlimited. The launch of Gemini 4 will be the moment of truth for this strategy.