Google has launched its new Gemini 3 Flash model, positioning it as a swift and cost-effective option built upon the Gemini 3 architecture introduced last month. This release aims to capture attention in a competitive landscape. The model is being set as the default within the Gemini app and for AI-powered search features.

Arriving six months after the Gemini 2.5 Flash, this latest version delivers notable advancements. In benchmark testing, Gemini 3 Flash shows a substantial leap over its predecessor and rivals the capabilities of other leading models, such as Gemini 3 Pro and GPT-5.2, in several areas.

For example, on the Humanity’s Last Exam benchmark, which assesses broad domain expertise, Gemini 3 Flash achieved 33.7% without tool assistance. This compares to scores of 37.5% for Gemini 3 Pro, 11% for Gemini 2.5 Flash, and 34.5% for the recently released GPT-5.2. On the MMMU-Pro benchmark for multimodality and reasoning, the new model led all competitors with a score of 81.2%.

**Consumer Rollout**

Globally, Google is replacing Gemini 2.5 Flash with Gemini 3 Flash as the default model in the Gemini app. Users can still select the Pro model from the settings for tasks involving math and coding.

The company highlights the new model's improved ability to interpret multimodal content and provide relevant answers. Practical applications include uploading a short pickleball video for tips, drawing a sketch for the AI to identify, or submitting an audio recording for analysis or quiz generation. Google also states the model better understands query intent and can produce more visual responses, incorporating elements like images and tables. Additionally, it can be used to create app prototypes directly within the Gemini app through prompts.

The Gemini 3 Pro model is now available to all users in the U.S. for search, and more U.S. users are gaining access to the Nano Banana Pro image model within search as well.

**Enterprise and Developer Availability**

Google reports that companies including JetBrains, Figma, Cursor, Harvey, and Latitude are already utilizing the Gemini 3 Flash model, accessible via Vertex AI and Gemini Enterprise. For developers, the model is available in a preview state through the API and within Antigravity, Google's new coding tool launched last month.

On the SWE-bench verified coding benchmark, Gemini 3 Pro scored 78%, a figure reportedly only exceeded by GPT-5.2. Google suggests the model is well-suited for video analysis, data extraction, and visual Q&A, and its speed makes it ideal for rapid, repetitive workflows.

**Pricing and Efficiency**

Pricing for Gemini 3 Flash is set at $0.50 per 1 million input tokens and $3.00 per 1 million output tokens. This represents a slight increase from the Gemini Flash 2.5 rates of $0.30 and $2.50, respectively.

However, Google asserts that the new model surpasses the performance of the Gemini 2.5 Pro while operating three times faster. For reasoning tasks, it uses an average of 30% fewer tokens than the 2.5 Pro, which could lead to overall cost savings for specific applications.

“We really position Flash as more of your workhorse model. So if you look at, for example, even the input and output prices at the top of this table, Flash is just a much cheaper offering from an input and output price perspective.”

The launch occurs amid heightened industry competition. Earlier this month, reports indicated OpenAI's Sam Altman issued an internal “Code Red” memo following a dip in ChatGPT's traffic as Google's consumer market share grew. Subsequently, OpenAI released GPT-5.2 and a new image generation model, while also highlighting expanded enterprise use and an 8x growth in ChatGPT message volume since November 2024.

While not directly addressing OpenAI, Google commented on the dynamic market environment. “Just about what’s happening across the industry is like all of these models are continuing to be awesome, challenge each other, push the frontier. And I think what’s also awesome is as companies are releasing these models,” Doshi said. “We’re also introducing new benchmarks and new ways of evaluating these models. And so that’s also encouraging us.”