Google Steals AI Lead

In the fast-moving world of AI language models (LLMs), the top players are shifting. Earlier, OpenAI and Meta were in control, but recent news shows Google’s latest AI is now getting the most attention.

Google was actually a bit behind, even though their own ideas led to the main tech used in today’s powerful AI. They even had a rough start with their Bard chatbot.

But lately, Google has put out strong new LLMs, and Meta and OpenAI have had some issues, which is changing the game.

Meta’s Llama 4 Launch Had Issues

Meta recently surprised everyone by releasing its new open-source LLM, Llama 4. The Saturday timing was unexpected for many.

Llama 4 can work with different kinds of information like pictures and sounds. It comes in a few sizes, including one called Scout that can remember a huge amount of text (up to 10 million tokens). This big memory helps it work with large documents or long conversations.

However, people weren’t entirely happy. Critics found out that the version of Llama 4 Meta used to get high scores on a popular ranking website (LMArena) was specially tweaked and not the standard one given to everyone. Meta also got criticized for saying Llama 4 Scout was great with its large memory, even though tests showed it wasn’t as good as other models at using very long texts. Plus, they didn’t release a model focused on complex “thinking” tasks right away.

Experts felt Meta might have rushed the announcement just to show they had a new model, even if it wasn’t fully ready.

OpenAI’s GPT-4.5 Was Too Expensive

OpenAI has also had setbacks. Their GPT-4.5 model was launched as their “biggest and best” for chat and did well in performance tests.

But the main problem was the price. Using GPT-4.5 through their developer tool (API) cost a massive $150 for every million output tokens. This was 15 times more than their GPT-4o model.

An AI expert mentioned that running such a huge model is hard with current technology and difficult to offer widely.

So, OpenAI announced they would stop offering GPT-4.5 through the API after less than three months. People can still use it through the ChatGPT website interface.

At the same time, OpenAI released GPT-4.1, a cheaper model ($8 per million tokens) that performs slightly less well overall but is better at some coding tasks. They also introduced new, more expensive models designed for reasoning.

Google Rises as Others Struggle

Llama 4 and ChatGPT-4.5 not being perfect created a chance for others, and Google’s new models jumped on it.

Some newer open-source models from Google (like Gemma) and other companies are now seen as better options than Llama 4 on leaderboards. They are strong, affordable to use, and some can even run on typical home computers.

But Google’s top-tier model, Gemini 2.5 Pro, made the biggest impression.

Launched in March, Gemini 2.5 Pro is built to “think” step-by-step. It understands different types of data, has a one-million-token memory, and is good at complex research.

Gemini 2.5 quickly moved up in rankings, winning some tests and now being the top model on LMArena. Google models generally hold many of the top spots there now.

Besides being powerful, Google is also competitive on price. Gemini 2.5 is free through their own apps and website, and using its API is reasonably priced. An even faster, cheaper version called Gemini 2.0 Flash is available for just 40 cents per million tokens.

Industry experts are noticing the change. One expert noted they use Google Gemini or other open models for complex thinking tasks because of OpenAI’s higher costs.

While Meta and OpenAI are still major players (ChatGPT has a billion users!), Gemini’s strong performance and good pricing show that the AI model race is definitely heating up, and Google is currently in a strong position.

This story was originally featured on IEEE Spectrum.