Google’s New Gemini 3 Flash Rivals Frontier Models At A Fraction Of The Cost

There’s nary extremity successful show for nan existent LLM merchandise cycle. Within nan past 30 days, we’ve seen nan launches of Google’s Gemini 3 Pro, Anthropic’s Opus 4.5, and OpenAI’s GPT-5.2. That’s successful summation to models from A2AI, DeepSeek, Grok, Mistral, Nvidia and others. Today it’s Google’s move again, pinch nan motorboat of nan smaller and faster type of Gemini 3: Gemini 3 Flash.

As we’ve seen pinch galore of nan smaller models from Google and different frontier exemplary builders, Gemini 3 Flash isn’t acold down its Pro brethren successful position of capabilities, pinch Gemini 3 Flash (with its reasoning mode on) being adjacent to Gemini 3 Pro, Anthropic’s Sonnet 4.5 and OpenAI’s GPT-5.2 successful astir benchmarks — and sometimes moreover beating them. Like its predecessor, it besides offers a 1 cardinal token discourse window.

To put Gemini 3 Flash’s capacity into perspective, conscionable a fewer weeks ago, Flash 3 would’ve been astatine nan apical of astir of nan frontier exemplary benchmarks. 

“For excessively long, AI forced a choice: large models that were slow and expensive, aliases high-speed models that were little capable. Gemini 3 Flash ends this compromise. Gemini 3 Flash delivers smarts and speed,” Google writes successful today’s announcement.

Compared to nan past Flash exemplary (Gemini 2.5 Flash), Gemini 3 Flash represents a important measurement up, which is particularly important for developers, arsenic Flash has agelong been recognized arsenic nan exemplary pinch nan champion price-to-performance ratio.

One area wherever Google has been particularly class-leading has been multimodal reasoning, pinch its models being capable to logic complete text, images, audio files and video. More recently, nan Gemini models person besides go rather tin astatine building their ain visualizations connected nan fly, thing Google besides highlights for this caller model. Indeed, Gemini 3 Flash moreover thumps Gemini 3 Pro successful nan multimodal MMMU-Pro benchmark, though only by 0.2%, 

Another area wherever Google’s models person precocious made immoderate advances is coding. On nan SWE-Bench Verified benchmark, Gemini 3 Flash besides thumps Gemini 3 Pro and is moreover up of Sonnet 4.5 (though GPT-5.2 remains nan apical performer here). 

“Gemini 3 Flash remains nan champion fresh for Warp’s Suggested Code Diffs, wherever debased latency and costs ratio are difficult constraints,” said Zach Lloyd, nan laminitis and CEO of Warp. “With this release, it resolves a broader group of communal command-line errors while staying accelerated and economical. In our soul evaluations, we’ve seen an 8% assistance successful hole accuracy.” 

One inclination we’ve precocious seen is that moreover these smaller models person been getting much costly for developers to usage done nan API, pinch Gemini 3 Flash now costing $0.5/$3 per cardinal input/output tokens, up from $0.3/$2.5. That’s still overmuch cheaper than Anthropic’s Claude Sonnet ($3/$5) aliases moreover nan smaller and little tin Claude Haiku ($1/5) models.

On average, though, Gemini 3 Flash uses 30% less tokens to make its answers erstwhile compared to Gemini 2.5 Flash, Google says, each while besides being faster. Google only compared this caller exemplary to nan older 2.5 Pro model, though, wherever it’s 3x faster.

The caller exemplary is now disposable successful nan API done Google AI Studio and Vertex AI, arsenic good arsenic successful nan company’s caller AI coding devices Antigravity, Gemini CLI and Android Studio. Google’s partners will besides build it into their ain tools, of course. 

For consumers, Gemini 3 Flash will now powerfulness Google Search’s AI Mode (with nan Pro exemplary still being an option, too) and nan “Fast” and “Thinking” modes successful nan Gemini app (where nan Pro mode will still beryllium available, too). 

Group Created pinch Sketch.