The new Flash model arrives six months after Google announced the Gemini 2.5 Flash model, offering significant improvements. On the benchmark, the Gemini 3 Flash model outperforms its predecessor by a significant margin and matches the performance of other frontier models, like Gemini 3 Pro and GPT 5.2, in some measures.

For instance, it scored 33.7% without tool use on Humanity’s Last Exam benchmark, which is designed to test expertise across different domains. In comparison, Gemini 3 Pro scored 37.5%, Gemini 2.5 Flash scored 11%, and the newly released GPT-5.2 scored 34.5%.

On the multimodality and reasoning benchmark MMMU-Pro, the new model outscored all competitors with an 81.2% score.

Consumer rollout

Google is making Gemini 3 Flash the default model in the Gemini app globally, replacing Gemini 2.5 Flash. Users can still choose the Pro model from the model picker for math and coding questions.

The company says the new model is good at identifying multimodal content and giving you an answer based on that. For instance, you can upload your pickleball short video and ask for tips; you can try drawing a sketch and have the model guess what you are drawing; or you can upload an audio recording to get analysis or generate a quiz.

The company also said the model better understands the intent of users’ queries and can generate more visual answers with elements like images and tables.

You can also use the new model to create app prototypes in the Gemini app using prompts.

The Gemini 3 Pro is now available to everyone in the U.S. for search and more people in the U.S. can access the Nano Banana Pro image model in search, as well.

Enterprise and developer availability

Google noted that companies like JetBrains, Figma, Cursor, Harvey, and Latitude are already using the Gemini 3 Flash model, which is available through Vertex AI and Gemini Enterprise.

For developers, the company is making the model available in a preview model through the API and in Antigravity, Google’s new coding tool released last month.

The company said the Gemini 3 Pro scores 78% on the SWE-bench verified coding benchmark, only outperformed by GPT-5.2. It added that the model is ideal for video analysis, data extraction, and visual Q&A, and because of its speed, it is suited for quick and repeatable workflows.

Image Credits:Google

Model pricing is $0.50 per 1 million input tokens and $3.00 per 1 million output tokens. This is slightly more expensive than $0.30 per 1 million input tokens and $2.50 per 1 million output tokens of Gemini Flash 2.5. But Google claims that the new model outperforms the Gemini 2.5 Pro model while being three times faster. And, for thinking tasks, it uses 30% fewer tokens on average than 2.5 Pro. That means overall, you might save on the number of tokens for certain tasks.

Image Credits:Google

“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. And so it actually allows for, for many companies, bulk tasks,” Tulsee Doshi, senior director & head of Product for Gemini Models, told TechCrunch in a briefing

Since it released Gemini 3, Google has processed over 1 trillion tokens per day on its API, amid its fierce release and performance war with OpenAI.

Earlier this month, Sam Altman reportedly sent an internal “Code Red” memo to the OpenAI team after ChatGPT’s traffic dipped as Google’s market share in consumers rose. Post that, OpenAI has released GPT-5.2 and a new image generation model. OpenAI also boasted about its growing enterprise use and said the ChatGPT messages volume has grown 8x since November 2024.

While Google didn’t directly address the competition with OpenAI, it said that the release of new models is challenging all companies to be active.

“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.”