Google расширяет возможности Gemini 3 Flash с мультимодальными функциями и повышенной производительностью Translation: Google Expands Gemini 3 Flash Capabilities with Multimodal Features and Enhanced Performance

Google has introduced the Gemini 3 Flash language model, which is now the default option in the Gemini app and the AI mode of its search engine.

Testing indicates that this new «lightweight» version of the LLM surpasses its predecessor, Gemini 2.5 Flash, and is competitive with top-tier models such as Gemini 3 Pro and GPT 5.2 in some performance metrics.

In the Humanity’s Last Exam, which gauges knowledge across various domains, Gemini 3 Flash scored 33.7%. For comparison, Gemini 3 Pro achieved 37.5%, Gemini 2.5 Flash scored 11%, and GPT-5.2 earned 34.5%.

In the MMMU-Pro benchmark, which assesses multimodal capabilities and reasoning skills, the neural network achieved a score of 81.2%, outperforming all competitors.

In the SWE-bench test, Gemini 2.5 Flash recorded 78%, trailing behind GPT-5.2. Google noted that the model is «perfectly suited» for video analysis, data extraction, and visual question answering.

Google has adopted Gemini 3 Flash as the standard model within the Gemini application, replacing Gemini 2.5 Flash. Users still have the option to select the Pro version for mathematics and programming inquiries.

According to the company, the new LLM effectively recognizes multimodal content and provides apt responses. Users can upload a short pickleball video and request advice, for instance.

The neural network has improved in understanding user intent and can generate more comprehensive responses using interactive elements like images and tables.

The cost for Gemini 3 Flash is set at $0.5 per million input tokens and $3 per million output tokens. For Gemini 2.5 Flash, the figures were $0.3 and $2.5, respectively.

However, the new model uses approximately 30% fewer tokens compared to 2.5 Pro for tasks that require cognitive effort, making it more cost-effective in certain scenarios.

Google’s tool for «vibe coding» and mini-application creation, Opal, is now available in Gemini.

Gems are customizable versions of Gemini introduced in 2024, designed for specific tasks or scenarios. These can serve as a coaching assistant, brainstorming partner, career guide, or coding collaborator.

Opal assists users in creating their mini-applications or combining existing ones simply by describing the desired features in natural language.

The visual editor in Opal outlines the necessary steps for building the tool.

OpenAI’s leadership in the AI race has been challenged by advancements from competitors, particularly Google.

In November, the company unveiled its flagship model, Gemini 3, deemed «the best in the world.» Subsequently, they released a new version of their image generation tool, Nano Banana Pro.

Gemini is increasingly becoming a popular solution and is closing the gap with ChatGPT in terms of popularity.

OpenAI’s chatbot continues to lead the AI segment, accounting for 50% of global mobile app downloads and 55% of monthly active users.

However, Gemini is outpacing its competitor in download trends, audience growth, and user engagement within the app.

This positive momentum may reduce the gap with ChatGPT. OpenAI has expressed concern, evidenced by a recent «red code» memo.

In response, OpenAI has expedited the releases of GPT-5.2 and a new version of the image generation tool, ChatGPT Images — GPT Image 1.5.

Recall that in December, Google introduced Disco, a browser tool based on Gemini 3, which transforms tabs into interactive applications.