Google’s Gemini 3 Flash: Smarter, Faster AI
▼ Summary
– Google has launched Gemini 3 Flash, a new, faster AI model now available in its app, search, and developer platforms.
– Benchmarks show Gemini 3 Flash outperforms its predecessor in academic and reasoning tests and has tripled its score in advanced knowledge exams.
– The model shows significantly improved coding skills, gaining nearly 20 points on the SWE-Bench Verified test compared to the last Flash version.
– Gemini 3 Flash is much more accurate on general-knowledge questions and performs closer to the Pro model while being more efficient.
– It retains the Pro model’s ability to generate interactive content, runs workloads three times faster than Gemini 2.5 Pro, and is cheaper than Pro models but more expensive than the previous Flash.
Google’s latest AI model, Gemini 3 Flash, represents a significant leap forward in balancing speed, capability, and cost for developers and users. This new iteration arrives as part of the broader Gemini 3 rollout, bringing enhanced performance directly to the Gemini app, Google Search, and various developer platforms like the Gemini API and Vertex AI. It is designed to be a smarter and faster alternative to its predecessor, offering improvements that narrow the performance gap with the more advanced Pro models.
The company emphasizes that Gemini 3 Flash is both faster and more capable than the previous Gemini 2.5 Flash model. Benchmark results support this claim, showing notable gains across several key evaluations. In foundational academic and reasoning tests such as GPQA Diamond and MMMU Pro, the new model outperforms the old one. Impressively, it even surpasses the Gemini 3 Pro model in the MMMU Pro benchmark. Its most dramatic improvement is visible in Humanity’s Last Exam (HLE), a test of advanced, domain-specific knowledge. Here, Gemini 3 Flash tripled the previous model’s score, achieving 33.7% without using any external tools, placing it just a few percentage points behind the Pro version.
A particular area of focus for this release is enhanced coding proficiency. Historically, Google positioned its Pro models as the superior choice for code generation. However, Gemini 3 Flash has made substantial progress, closing much of that gap. On the SWE-Bench Verified test, a popular benchmark for evaluating coding ability, the new model gained nearly 20 points over the earlier 2.5 Flash branch, indicating a major step forward in understanding and generating functional code.
Accuracy on general knowledge questions has also seen a remarkable boost. In the Simple QA Verified test, Gemini 3 Flash scored 68.7%, a figure that approaches the performance of the Pro model and is a dramatic increase from the previous Flash model’s score of just 28.1%. These evaluation scores suggest that the performance profile of Gemini 3 Flash is now much closer to that of the Pro-tier models from the older generation, while maintaining greater operational efficiency.
The model inherits a key advancement from Gemini 3 Pro: the underlying capability to generate interactive simulations and multimodal content. This means developers can access sophisticated content creation features without necessarily requiring the top-tier model. Crucially, Gemini 3 Flash delivers better performance than the older Gemini 2.5 Pro while processing workloads three times faster.
From a cost perspective, it remains a more budget-friendly option than the Pro models, though pricing has increased from the previous Flash generation. Developers will pay $0.50 per million input tokens and $3 per million output tokens for Gemini 3 Flash. This compares to the previous rates of $0.30 and $2.50 for Gemini 2.5 Flash, and the Pro model’s costs of $2 and $12 for input and output tokens, respectively. This pricing structure positions it as a powerful and cost-effective solution for scalable AI applications.
(Source: Ars Technica)
