Can the Gemini 3 truly outperform the GPT-5.1?

To be honest, when I saw the Gemini 3 Pro data set, my first reaction was: isn't this really a test data leak?With a 1501 Elo score and a 45.81 TP3T human final exam score, these numbers are truly shocking. But calm down and think about it, does a high score on a benchmark test mean true transcendence? After all, GPT-5.1 has accumulated quite a good reputation for its performance in real application scenarios, such as code generation and creative writing, which are the areas we touch most on a daily basis.

Behind the Performance Data

While the review scores are certainly eye-catching, it's the actual experience of using it that counts. One interesting detail I noticed: the Gemini 3 did perform amazingly well in the multimodal tests, especially in video comprehension with a score of 87.61 TP3T, which is really a cut above the GPT-5.1. Then again, these tests are conducted on specific datasets, and real-world application scenarios tend to be much more complex. Take my own experience, sometimes the model performs excellently on the test set, but when dealing with some sudden, unstructured tasks, it still comes up with situations that make you cry.

And did you guys notice that Gemini 3's contextual window of a million tokens is really a killer. Imagine dropping the entire book "Three Bodies" in for the AI to analyze, and it can actually make sense of the ambiguities and character relationships - an ability that really hasn't been seen on GPT-5.1 yet. But then again, how valuable is such a large context window in practical applications? After all, we don't normally use AI and most of the time we don't need to deal with such long texts.

The Practical Value of Deep Thinking Models

That case of being able to write poems explaining plasma physics is really cool, but I'm wondering if that ability is really that important to the average user. What interests me more is its performance in programming - creating 3D spaceship games and also being able to add sound effects, this is really better than the current GPT-5.1. To be honest, though, I found in real-world testing that while Gemini 3 performs well on complex tasks, it still has occasional comprehension biases on some basic natural language understanding tasks.

Another factor that has to be taken into consideration is the maturity of the ecosystem; the GPT series has developed a well-established developer ecosystem and toolchain after so many years of iterations, and Gemini 3, despite its impressive technical specifications, will take a while to build up an equally mature ecosystem. What do you guys think?

At the end of the day, the competition of AI models is never just about technical parameters, but user experience, responsiveness and price are the key to determine whether a model can really win users. User experience, response speed, price factor, these seemingly "soft" indicators, is often the key to determine whether a model can really win users.Gemini 3 does show a breakthrough in some aspects, but to say that the full range of beyond, it may be too early. After all, this AI race is like a marathon, the current lead does not mean victory at the finish line.