Google has just released its latest voice model, Gemini 2.5 Pro, for free users. The new model, which is still in an ‘experimental’ phase, can be selected from the drop-down menu on the Gemini website, while support for the iOS and Android mobile apps is expected soon. Google had announced the Gemini 2.5 Pro model earlier this week for advanced Gemini users, but with ChatGPT hogging all the limelight with its native image generation feature that powers the viral Ghibli image trend, the tech giant has decided to give the update to free users as well.
Gemini 2.5 Pro is a reasoning model, just like OpenAI’s o3 Mini or DeepSeek R1. Unlike pre-trained models or GPTs, reasoning models aim to mimic human-like reasoning abilities to achieve better performance and accuracy.
What’s new with Gemini 2.5 Pro?
Google claims that Gemini 2.5 Pro is strong at reasoning and coding capabilities-related tasks. The latest language model also leads in common coding, math and science-related tasks in benchmarks such as Humanity’s Last Exam and LMArena.
While Google has traditionally lacked behind OpenAI and Claude in coding capabilties, the company has bet big on this area with Gemini 2.5 Pro by adding a massive context window of 1 million tokens. The context window is the amount of text that the chatbot can process in one go with a 1 million context window, roughly translating to 7,50,000 words.
In comparison, Claude’s latest 3.7 Sonnet model has a context window of 5,00,000 tokens while OpenAI’s o3 Mini has a 2,00,000 context window.
In its demo video, Google showed Gemini 2.5 Pro creating a game using a single line of text prompt. Again, Google refers to a number of benchmarks to make its point about the programming capabilities of Gemini 2.5 Pro, but these benchmarks change quickly as new models are released and actual performance is likely to depend on the end user’s use case.
Google also added that it is building these capabilities directly into all of its models to handle more complex problems and support context-aware agents.
AI race at a boiling point:
OpenAI kicked off the AI race with the launch of its GPT-3.5 model in late 2022. Since then, almost every major tech company has had a razor-sharp focus on developing their own AI offerings, from Microsoft to Google to Meta.
The tech giants are now seemingly moving away from pre-trained models and focusing more on reasoning models, which are also used to power AI agents such as deep search.
While US-based AI companies were thought to be ahead of the curve until recently, China’s DeepSeek broke all those notions with its DeepSeek R1 and DeepSeek V3 models, which were not only built at a fraction of the cost (despite import restrictions on AI chips), but also rivalled the competency of leading Western AI models.
At the height of DeepSeek’s popularity earlier this year, panic swept through the US stock market and millions were wiped out. Since then, US tech companies have been rushing to launch newer and newer models to underline their lead in the AI race, while DeepSeek is also reportedly planning to launch a new model soon.
