Google is no longer chasing the same finish line as OpenAI and Anthropic in the AI race. While competitors pursue recursive self-improvement, a concept where AI systems build better versions of themselves, Google DeepMind is betting heavily on world models, artificial intelligence that learns how the physical world actually works. The strategic divergence is subtle but significant, and it's reshaping how we should think about who's winning the AI competition.

What Exactly Is Google Betting On Instead of Coding Agents?

A world model is fundamentally different from the large language models most people interact with today. Instead of predicting the next word in a sentence, world models learn how gravity works, how objects move, and what happens when you push something. They simulate cause and effect in the real world. DeepMind's recent releases, including Genie 3 and SIMA 2, an agent that learns by playing inside virtual 3D worlds, signal this shift. The lab even extended Project Genie to Google Street View, suggesting the company is serious about building AI that understands physical spaces.

This contrasts sharply with what OpenAI and Anthropic are pursuing. Both companies are focused on recursive self-improvement, or RSI, which means creating AI systems capable of building the next generation of AI without human intervention. Anthropic has already documented progress here; the company reported that Claude wrote more than 80% of the code it ships by May 2026, up from near zero before February 2025.

Is Google Actually Losing Ground in the AI Race?

Google's latest model release, Gemini 3.6 Flash, landed 10th on one independent ranking, a significant drop from where the company typically sits. The pitch for the new model was speed and cost efficiency, not raw power. Gemini 3.6 Flash produces 17% fewer tokens, the small chunks of text an AI writes, than its predecessor, making it cheaper to run but not necessarily more capable.

However, the story is more nuanced than a simple decline. Google still leads on one crucial benchmark: MLE-Bench, which tests whether a model can build machine-learning systems on its own. A Gemini 3 model scored 64.4% in February, the best result at the time, and Gemini 3.6 Flash achieved 63.9% in July. Additionally, Google's Gemini app has 950 million monthly users, demonstrating the company's continued market dominance.

Sundar Pichai, Google's CEO, has explicitly tied the company's roadmap to personalized AI agents rather than leaderboard rankings. This suggests the company is making a deliberate choice about where to focus its efforts, not simply falling behind.

Why Would Google Deliberately Step Back From the Self-Improvement Race?

One compelling outside perspective came from Jack Clark, co-founder of Anthropic. In May 2026, Clark published an analysis assessing which labs are pursuing recursive self-improvement. He gave a 60% probability that AI could run its own research by the end of 2028, with a 30% chance by 2027. When evaluating which major labs are chasing this goal, Clark described DeepMind as "the most circumspect of the big three," meaning the most cautious.

Clark, co-founder of Anthropic

"Hassabis is betting on something else: world models. Models that can understand and simulate the real world, not just predict the next token," wrote Alberto Romero in a recent analysis.

Alberto Romero, AI analyst

Clark's evidence came from DeepMind's own 2025 safety paper, co-written by co-founder Shane Legg, which suggested the lab is taking a more cautious approach to AI advancement. By contrast, Anthropic has been transparent about its progress on self-improvement, reporting a 52-fold gain in one speed test in April 2026, compared to a 2.9-fold gain a year earlier.

How to Understand Google's Long-Term AI Strategy

• World Models Focus: Google is investing in AI that understands physical reality and can simulate real-world scenarios, positioning itself for applications in robotics and embodied AI rather than pure language tasks.
• Cost and Efficiency Over Raw Power: Recent releases like Gemini 3.6 Flash prioritize speed and affordability, suggesting Google is optimizing for practical deployment at scale rather than benchmark dominance.
• Patience Backed by Search Revenue: Google's search business generates $63.3 billion in quarterly revenue, giving DeepMind financial runway to pursue longer-term bets without immediate pressure to match competitors on leaderboards.

What's the Financial Reality Behind Google's AI Spending?

Google's patience comes with a price tag. Alphabet poured $44.9 billion into data centers and equipment in the June quarter alone, roughly double the amount from a year earlier. This aggressive spending has created a cash flow problem: the company reported negative free cash flow of $5.86 billion in the quarter ending June 2026, compared to positive $10.1 billion in March and positive $24.6 billion in December.

To cover the gap, Alphabet sold $49.6 billion in new shares and borrowed another $20.3 billion. Long-term debt doubled in six months, from $46.5 billion to $98.2 billion. A line item in Alphabet's accounts covering shared AI research lost $5.79 billion, up from $3.37 billion.

This financial pressure means Google's strategy will face real scrutiny. The company can afford to pursue world models for now, but the next earnings report will reveal how long Alphabet can sustain this spending without returning to profitability.

What Should We Watch Next?

Several milestones will determine whether Google's world-model bet was prescient or a strategic misstep. The release of Gemini 3.5 Pro, currently in testing with partners, will show whether Google can still compete on raw capability. The arrival of Gemini 4, which is undergoing Google's biggest training run yet, will be the real test. If world models deliver breakthroughs where coding agents have stalled, Google's slower pace will look strategic rather than defensive.

Demis Hassabis, who runs Google DeepMind, has never publicly ruled out pursuing recursive self-improvement, leaving room for the company to shift direction if world models don't deliver. For now, Google is betting that understanding the real world matters more than building AI that builds itself. Whether that bet pays off will reshape not just Google's future, but the entire trajectory of artificial intelligence development.