Gemini 3 Flash is here with reduced cost and latency – a powerful combo for the enterprise

Thanks to the newly released Gemini 3 Flash, enterprises can now harness the power of a large language model that approaches the state-of-the-art of Google’s Gemini 3 Pro, at a fraction of the cost, and at a fraction of the cost.

This model joins the flagship Gemini 3 Pro, Gemini 3 Deep Think, and Gemini Agent that were announced and released last month.

Gemini 3 Flash is now available in preview for Gemini Enterprise, Google Antigravity, Gemini CLI, AI Studio, and Vertex AI to process information in near real-time and help you build fast, responsive agent applications.

Gemini 3 Flash is “built on a series of models that developers and enterprises already love, and is optimized for high-frequency workflows that demand speed without sacrificing quality,” the company said in a blog post.

This model is also the default for AI mode in Google Search and Gemini applications.

Tulsee Doshi, senior director of product management for the Gemini team, said in a separate blog post that this model “shows that speed and scale don’t have to come at the expense of intelligence.”

“Gemini 3 Flash is built for iterative development and delivers the pro-grade coding performance of Gemini 3 with low latency, allowing you to quickly reason and solve tasks in high-frequency workflows,” said Doshi. “It provides the ideal balance between agent coding, production-ready systems, and responsive, interactive applications.”

Early adoption by specialized companies has proven the reliability of this model in high-stakes fields. Harvey, an AI platform for law firms, reported a 7% increase in inference on its internal “BigLaw Bench.” Meanwhile, Resemble AI found that Gemini 3 Flash can process complex forensic data for deepfake detection four times faster than Gemini 2.5 Pro. These are not just speed improvements. It enables “near real-time” workflows that were previously not possible.

lower cost and more efficient

Enterprise AI builders have become more aware of the cost of running AI models, especially as they try to convince stakeholders to spend more on agent workflows running on expensive models. Organizations are turning to smaller and distilled models, focusing on open models and other research and promoting techniques to manage bloated AI costs.

For enterprises, Gemini 3 Flash’s biggest value proposition is that it offers the same level of advanced multimodal functionality as the larger Gemini Flash, such as complex video analysis and data extraction, but is much faster and cheaper.

Google’s internal documentation highlights a 3x speedup compared to the 2.5 Pro series, but data from independent benchmarking firm Artificial Analysis adds another important nuance.

In the latter organization’s pre-release testing, Gemini 3 Flash Preview recorded a raw throughput of 218 output tokens per second. This makes it 22% slower than the previous “non-inference” Gemini 2.5 Flash, but still significantly faster than rivals such as OpenAI’s GPT-5.1 High (125 t/s) and DeepSeek V3.2 Inference (30 t/s).

Most notably, Artificial Analysis recognized Gemini 3 Flash as a new leader in the AA-Omniscience Knowledge Benchmark, achieving the highest knowledge accuracy of any model tested to date. However, this intelligence comes with an “inference tax.” When working on complex indexes, this model uses more than twice the amount of tokens compared to the 2.5 Flash series.

This high token density is offset by Google’s aggressive pricing. When accessed via the Gemini API, Gemini 3 Flash costs $0.50 per million input tokens compared to $1.25 per million input tokens for Gemini 2.5 Pro, and output tokens cost $3 per million output tokens for Gemini 2.5 Pro. This allows Gemini 3 Flash to claim the title of the most cost-effective model in its intelligence tier, despite being one of the most “talkative” models in terms of raw token volume. Here’s how it compares to competing LLM products:

Other storage methods

However, enterprise developers and users can further reduce costs by eliminating the delays associated with increased token usage that are common in most large-scale models. Google says the model “can adjust how much it thinks,” so it uses more thought and uses more tokens for complex tasks than for quick prompts. The company noted that Gemini 3 Flash uses 30% fewer tokens than Gemini 2.5 Pro.

To balance this new reasoning ability with the stringent latency requirements of enterprises, Google introduced the “Level of Thought” parameter. Developers can toggle between “Low” to minimize cost and latency for simple chat tasks, and “High” to maximize inference depth for complex data extraction. This fine-grained control allows teams to build “variable-speed” applications that only spend expensive “thinking tokens” when the problem actually requires PhD-level ability.

The economic story goes beyond simple token prices. With built-in context caching, businesses working with large static datasets, such as entire legal libraries or codebase repositories, can reduce the cost of repetitive queries by 90%. Combined with the 50% discount on the Batch API, the total cost of ownership for Gemini-powered agents is well below the threshold of competing Frontier models.

“Gemini 3 Flash delivers great performance on coding and agent tasks at a low price, allowing teams to introduce advanced inference costs across processes at scale without hitting any roadblocks,” Google said.

Google claims that by offering a model that delivers strong multimodal performance at a more affordable price, businesses concerned about controlling their AI spending should choose its models, especially Gemini 3 Flash.

Strong benchmark performance

But how does Gemini 3 Flash compare to other models in terms of performance?

According to Doshi, the model achieved a score of 78% on Coding Agent’s SWE-Bench Verified benchmark test, outperforming both the previous Gemini 2.5 family and the new Gemini 3 Pro itself.

This means that companies can offload large amounts of software maintenance and bug fixing tasks to models that are faster and cheaper than previous flagship models without compromising code quality.

This model also performed well on other benchmarks, scoring 81.2% on the MMMU Pro benchmark, matching the Gemini 3 Pro.

While most Flash-type models are explicitly optimized for short, quick tasks such as code generation, Google says Gemini 3 Flash’s performance is “ideal for developers who want to perform more complex video analysis, data extraction, and visual Q&A in inference, tool usage, and multimodal functionality, including in-game assistants and A/B.” This means more intelligent applications, such as testing experiments, that require both quick answers and deep reasoning.

First impressions of early users

So far, early users have been very impressed with this model, especially its benchmark performance.

What it means to use AI in the enterprise

With Gemini 3 Flash now serving as the default engine for Google Search and the entire Gemini app, here’s what we’re seeing: "Flashing" Frontier Intelligence. Google is setting a trap for slow incumbents by making professional-level reasoning the new baseline.

Integration into platforms like Google Antigravity suggests that Google isn’t just selling models. We’re selling infrastructure for autonomous enterprises.

As developers hit the ground running with 3x faster speeds and 90% discounts on context caching. "gemini first" Strategy makes for a compelling financial argument. In the high-speed race for AI supremacy, Gemini 3 Flash may finally be the model to flip "vibe coding" From an experimental hobby to a production-ready reality.

Source link