I just wanted to introduce you to the latest AI model from Google, Gemini 2.5 Pro. This brand-new model is now available to all of us within our Google Workspace tools through the Gemini program. Rather than doing a tutorial as usual, I thought I’d quickly go over this benchmark chart. Think of the benchmarks like standardized tests or obstacle courses designed to push these AI brains to their limits. It shows how Gemini 2.5 Pro performs.




First, let's talk about Reasoning. One tough test is "Humanity's Last Exam." The exam focuses on solving complex puzzles and understanding tricky situations, going beyond just knowing facts Gemini 2.5 Pro scored 18.8% on this. While this might sound low, you can see it was a lot higher than all the other AIs in this comparison, showing strong capabilities in handling very complex problems.

How about Science and Math? These benchmarks are like advanced university-level exams. We see tests like "GPQA diamond" for science and "AIME" for competitive math problems. Here again, Gemini 2.5 Pro put up impressive scores, often getting the highest marks (like 84% in science, and 86.7% and 92% in two different math tests) when only given one try at each problem. This shows a strong grasp of these technical subjects compared to models like OpenAI's o3-mini and Claude 3.7 Sonnet in these specific tests.

What about Coding? Benchmarks like "LiveCodeBench" test writing new code, "Aider Polyglot" checks how well they edit existing code, and "SWE-bench" sees if they can tackle software engineering tasks. The results here are interesting!

While OpenAI's o3-mini showed a slight edge in writing new code on one benchmark, Gemini 2.5 Pro demonstrated very strong performance in editing code and handling more complex software engineering tasks, often outperforming models like Claude 3.7 Sonnet in those areas.

This model can also understand Images and Visuals. Tests like "MMMU" and "Vibe-Eval" measure this. Gemini 2.5 Pro scored highly here (81.7% and 69.4%, meaning it correctly interpreted images and visuals the vast majority of the time), demonstrating strong "eyesight," an area where some competitors currently lack capabilities.

It also showed it can handle really Long Documents (scoring 94.5% on the 'MRCR' test) and understand many Different Languages (scoring 89.8% on "Global MMLU"). Finally, on a Factuality test called "SimpleQA" – basically checking how often it tells the truth – Gemini 2.5 Pro did well (52.9%), though OpenAI's GPT-4.5 scored higher on this specific test.
Gemini 2.5 also, as you might have guessed, wrote most of this script after analyzing the table as a jpg.

So, what's the takeaway? These benchmarks show us that the new model, the one we can all use today in Workspace, is super versatile – like having a digital expert in many fields right at our fingertips. While different models have specific strengths, this data paints Gemini 2.5 Pro as the best AI model in the world right now, being a very well-rounded performer. I encourage you to check it out when you get the chance!