Google's AI has been upgraded again.

The Gemini 3 Flash model offers high speed and low operating costs, performing just as well as Gemini 3 Pro or GPT-5.2.

The Gemini 3 Flash model focuses on cost and processing efficiency. Photo: Bloomberg .

Google has just launched Gemini 3 Flash. Based on Gemini 3, this is a high-speed, low-cost AI model that competes with OpenAI's newer GPT versions.

Gemini 3 Flash was released six months after Google introduced Gemini 2.5 Flash. Benchmark reviews showed the new model outperformed its predecessor, matching advanced models like Gemini 3 Pro and GPT-5.2 in several categories.

For example, Google claims Gemini 3 Flash scored 33.7% on the Humanity's Last Exam test. For comparison, Gemini 3 Pro scored 37.5%, Gemini 2.5 Flash scored 11%, while the newly released GPT-5.2 scored 34.5%.

The new model also achieved 81.2% on the MMMU-Pro reasoning scale, which measures multimodal reasoning ability. This figure is higher than most competing models.

Early release for users

Google says Gemini 3 Flash will be the default model in the Gemini app, replacing version 2.5 Flash. Users can still switch to Gemini 3 Pro for more complex tasks, such as math or programming questions.

Gemini 3 Flash is also highly regarded for its multimodal content processing capabilities. For example, users can provide videos of pickleball gameplay and request technical analysis, sketches to predict patterns, or upload audio files for the tool to analyze and create questionnaires.

Gemini 3 Flash's ability to analyze video content. Image: Google .

"Thanks to Gemini 3 Flash's powerful multimodal reasoning capabilities, you can use the tool to receive, analyze, and process a wide variety of information at high speed."

For example, you can ask Gemini to analyze videos/images and transform the content into a detailed, actionable plan in seconds," Google's announcement emphasized.

When using AI Mode, Gemini 3 Flash is also able to better understand the intent behind commands, thereby providing visual responses with illustrations or charts. Users can program based on inspiration (vibe-coding) to create applications in Gemini thanks to the Opal tool.

To facilitate more efficient programming.

For developers, Google is offering Gemini 3 Flash as a beta version through Gemini AI in AI Studio and Antigravity, an agent-based programming tool.

Google highlighted Gemini 3 Flash's programming capabilities. This model achieved 78% on the SWE-bench Verified, a benchmark used to evaluate programming agents, higher than Gemini 3 Pro and only behind GPT-5.2.

"Gemini 3 Flash's powerful capabilities in inference, tooling, and multimodal processing make it an ideal choice for developers who want to perform complex tasks such as video analysis, data extraction, and image-based querying," Google said.

The cost of Gemini 3 Flash is optimized at $0.50 per 1 million input tokens and $30 per 1 million output tokens. If the input data uses audio, the cost remains at $1 per 1 million tokens.

This price is slightly higher than the Gemini 2.5 Flash ( 0.3 USD and 2.5 USD respectively). However, Google claims this model performs better than the Gemini 2.5 Pro, being three times faster.

The performance of Gemini 3 Flash across several metrics. Image: Google .

For inference tasks, Gemini 3 Flash uses 30% fewer tokens than 2.5 Pro, which theoretically saves users money on certain requirements.

Since the launch of Gemini 3, Google says it has processed over 1 trillion tokens per day on its API platform. Users commonly use Gemini 3 for vibe-coding tasks, interactive game design, and multimodal content analysis.

The tech community considers Gemini 3 a major competitor to GPT-5. In early December, OpenAI CEO Sam Altman reportedly sent a "red alert" to employees as ChatGPT traffic dropped. OpenAI quickly launched GPT-5.2 and a new image generation model.

While not directly mentioning OpenAI, Google stated that the launch of the new model requires all companies in the market to take proactive action.

"What's happening across the industry is that all the models are still good, competitive, and mutually reinforcing. Companies are actively releasing these models," Tulsee Doshi, Product Director of Gemini Models, told TechCrunch .

Source: https://znews.vn/ai-cua-google-lai-duoc-nang-cap-post1612582.html