Google says its latest reasoning model is its “most intelligent” — but Microsoft’s CEO claims Google already fumbled its AI opportunity

As an analyst, I’m thrilled to share that Google has unveiled the next generation of their AI model, Gemini 2.5. They’re touting it as their most intelligent yet. The initial version of this new model is none other than Gemini 2.5 Pro, which has left an indelible mark with its exceptional performance across a multitude of tests.

According to Google, their Gemini 2.5 model surpasses the top performances of OpenAI, DeepSeek, and leading AI models from other major technology companies.

The current version of Gemini 2.5 Pro can be accessed via Google AI Studio and the Gemini app for advanced users. In the upcoming period, it will also be accessible through Vertex AI.

At this time, Google has not shared pricing for Gemini 2.5 Pro or other Gemini 2.5 models.

Models that incorporate Gemini 2.5 function as “cognitive models.” This implies they engage in a series of mental operations or reasoning before producing an answer. These “thoughtful” models are gaining significant attention within the AI community due to their ability to generate more intricate responses and offer higher precision levels.

As I observe, Google has announced the launch of Gemini 2.5, which they claim has taken performance to a new height. This is made possible by merging an advanced baseline model with refinements in its post-training process.

In the future, we’ll be integrating problem-solving skills into all our models. This way, they can tackle intricate issues and create more competent, adaptable agents that understand their environment better.

Gemini 2.5 vs. OpenAI models

Google has shared some remarkable benchmark scores for Gemini 2.5, and the experimental version of Gemini 2.5 even achieved a score of 18.5% on the final exam known as Humanity’s Last Exam.

For the time being, it appears that Gemini 2.5 Pro Experimental outperforms both OpenAI 03-mini (by 14 percentage points) and DeepSeek R1 (by 5.9 percentage points), making it the top model according to this specific metric.

The particular test is generally perceived as challenging, but it’s important to note that it’s not the sole method for evaluating an AI model’s performance.

Google has emphasized that the Gemini 2.5 Pro excels in coding capabilities, and it stands out in math and science evaluations, particularly when assessed using GPQA and AIME 2025 benchmarks, making it a top performer in these fields.

Can you code with Gemini 2.5?

In Gemini 2.5, coding plays a significant role. Google boasts about making a substantial advancement from 2.0 and hints at further enhancements yet to be revealed.

The latest Google model is capable of generating web applications and programming for autonomous applications, as demonstrated by Google with a showcase of the experimental version, Gemini 2.5 Pro, being utilized to develop a game merely from a one-line command.

Read More

2025-03-26 19:39