Google is no longer racing OpenAI and Anthropic to the same finish line. Its rivals want AI that improves itself. Google wants AI that understands the real world.

The split is easy to miss, because Google still ships models and still makes money. But its newest release landed 10th on one independent ranking.

What Google Has and Has Not Said About the AI Race

Google released Gemini 3.6 Flash on July 21. The pitch was speed and cost, not raw power.

Follow us on X to get the latest news as it happens

The model produces 17% fewer tokens than the version before it, Google’s blog said. Tokens are the small chunks of text an AI writes. Fewer tokens means a cheaper answer.

Power is another matter. One published reading of the Artificial Analysis index placed the model 10th. Every other major lab ranked above it.

Google is not standing still. It has begun its biggest training run yet, for Gemini 4. A larger model, Gemini 3.5 Pro, is still in testing with partners.

Sundar Pichai has pointed to a different prize. He has tied the roadmap to personalized agents rather than leaderboard wins.

Investors are less relaxed. Alphabet shares fell 6% in June after two senior researchers left for rivals.

Inside DeepMind’s Bet on World Models

A world model is AI that learns how physical things behave. Gravity, motion, cause and effect. It predicts what happens next in a room, not the next word in a sentence.

DeepMind’s own website shows the bet. It files Genie 3 and Gemini Robotics under a heading for world models and embodied AI, meaning software that controls machines.

In May the lab extended Project Genie to Street View. It also released SIMA 2, an agent that learns by playing inside virtual 3D worlds.

OpenAI and Anthropic are aiming somewhere else entirely. They want recursive self-improvement, shortened to RSI.

In plain terms, that is AI clever enough to build the next, better AI. Then that one builds the one after it.

Writer Alberto Romero argued on Tuesday that Google left this race on purpose.

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

Google has said no such thing. Demis Hassabis, who runs Google DeepMind, has never ruled out RSI in public.

Why a Rival Co-Founder Says DeepMind Is the Outlier

The sharpest outside read came from a rival, months earlier.

Jack Clark co-founded Anthropic. On May 4 he published an essay on where AI is heading.

He gave a 60% chance that AI can run its own research by the end of 2028. He put 2027 at 30%.

Clark then asked which labs are chasing that goal. DeepMind, he wrote, “appears to be the most circumspect of the big three.” Circumspect means cautious.

His evidence came from DeepMind itself. He cited its 2025 paper on AI safety, co-written by co-founder Shane Legg.

Anthropic is far bolder. It reported in a recursive self-improvement study that Claude wrote more than 80% of the code it ships by May 2026.

Before February 2025, that share was near zero.

The firm has documented AI building better AI. On one speed test, its models delivered a 52-fold gain in April, against 2.9-fold a year earlier.

A skilled engineer needs four to eight hours to manage a fourfold gain on the same task.

Why Google Might Not Have Quit the AI Race at All

Two facts cut against the whole idea:

• Google leads the test that comes closest to measuring AI research skill.

MLE-Bench asks a model to build machine-learning systems on its own. A Gemini 3 model scored 64.4% in February. That was the best result at the time. Google put 3.6 Flash at 63.9% in July.

Labs that quit a field rarely top its scoreboard.

• Google is nowhere near absent from the market.

Pichai told investors the Gemini app has 950 million monthly users. Google did skip NVIDIA’s open AI alliance this month. So did OpenAI and Anthropic.

Can Google Afford to Wait?

The case for patience is simple. Search ads pay for everything else, so DeepMind can take its time.

Alphabet’s own filing shows that cushion getting thinner.

Revenue reached $119.8 billion in the June quarter, up 24%, according to results filed July 22. Search alone brought in $63.3 billion.

Then comes the spending. Alphabet poured $44.9 billion into data centers and equipment in three months. That is roughly double a year earlier.

The result was negative free cash flow of $5.86 billion. Free cash flow is what is left after the building bills are paid.

That figure was positive $10.1 billion in March. In December it was positive $24.6 billion.

Alphabet covered the gap by selling $49.6 billion of new shares in June. It borrowed another $20.3 billion.

Long-term debt doubled in six months, from $46.5 billion to $98.2 billion.

A line in the accounts covering shared AI research lost $5.79 billion, up from $3.37 billion. Patience now carries a price tag.

What to Watch Over the Next 30 Days

• Whether Gemini 3.5 Pro ships, and how it ranks
• Whether DeepMind shows world-model results tied to Gemini 4
• Whether Alphabet cash flow turns positive again in September
• Whether Hassabis answers the self-improvement question directly

Gemini 4 is the real test. If world models work where coding agents stall, the slow pace will look smart rather than scared.

The next earnings report will show how long Alphabet can keep paying to find out.

Source: BeInCrypto