Google has pushed out a shiny new AI model in the form of Gemini 2.5 Pro, albeit with an experimental label next to it—and it’s available for free, so you don’t need to subscribe to Gemini Advanced to get it. As with many recent AI model releases, the “reasoning” capabilities of the model are said to be the biggest upgrade here.
In artificial intelligence terms, reasoning means answers that are more thoroughly worked through. That should produce fewer mistakes, more logical responses, and a better appreciation of “context and nuance” according to Google. This capability for extra “thought” will now come as standard in future Google models.
The Pro (Experimental) release is the first variant of Gemini 2.5 to show up, and while the original blog post didn’t mention free users, less than a week later we’ve got an update saying it’s available for everyone—with rate limits applied if you’re not a Gemini Advanced subscriber (Google hasn’t specified what those rate limits are). The new model is available now through the desktop app, and coming soon to mobile.
Credit: Google
Google points to several benchmark tests that show the prowess of Gemini 2.5 Pro. At the time of writing it tops the LMArena leaderboard, where users give ratings on responses from dozens of AI chatbots. It also scores 18.8 percent on the Humanity’s Last Exam test—which measures human knowledge and reasoning—narrowly edging out rival models from OpenAI and Anthropic.
Also of note: the large context window. In simple terms, this is an indicator of how much data the AI model can churn through in one go, and Gemini 2.5 Pro has a context window of one million tokens, with two million “coming soon” according to Google. That compares to a context window of, for example, 200,000 tokens for ChatGPT’s o3-mini reasoning model.
As tends to be the norm with these AI announcements, there’s no mention of copyright infringement as far as training data goes, or increasing energy use. According to MIT researchers, modern-day AI models use a “staggering” amount of electricity and water, and have put us on an “unsustainable path” that needs to change direction quickly.
Putting Gemini 2.5 Pro to the test
It can be tricky to quantify improvements from one AI model to the next, which is why benchmarks like LMArena are useful. I lack the expert scientific or programming knowledge needed to really put Gemini 2.5 Pro to the test—though as with the previous model, I was able to create some simple web apps (like an online timer) in minutes.
I do know a bit about Charles Dickens’ Bleak House, so I set Gemini 2.5 Pro to work on the text. It gave me an accurate summary of the plot, and a clever assessment of the different narrative devices used (which would’ve really helped me in my study days). It also converted the book into a reasonably well done three-act structure for a movie—evidence of it holding a lot in its “mind” at once.
The older Gemini 2.0 Flash was able to answer the same Bleak House prompts accurately too, but the responses from Gemini 2.5 Pro were longer, more detailed, less generic, and smarter—evidence of that extra “reasoning” being put to work. The Gemini 2.0 Flash model also had to split the movie adaptation into three responses, perhaps due to the sheer amount of text it was trying to process.
Google has provided its own example of the capabilities of Gemini 2.5 Pro, showing how a simple endless runner game can be produced from a single prompt. While the demo video showing the code output is sped up, the game does appear to work and be pretty well-designed, which is an impressive end result from a single natural language prompt. There’s also a neat web demo of digital fish swimming around.
Elsewhere on the web, the new AI model is being extensively tested. Software engineer and independent AI researcher Simon Willison ran several tests covering image creation, audio transcription, and code generation, and came away very much liking what Gemini 2.5 Pro had been able to come up with.
The frenetic pace of AI development shows no signs of slowing down anytime soon, and we can expect more Gemini 2.5 models to appear in the near future. “As always, we welcome feedback so we can continue to improve Gemini’s impressive new abilities at a rapid pace, all with the goal of making our AI more helpful,” says Koray Kavukcuoglu, from Google’s DeepMind AI lab.
