The release of Gemini’s new AI models on July 21, 2026, marks a significant leap forward in the efficiency and performance of AI language models. Featuring Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber, the evolution of these models is apparent through their enhanced token usage efficiency and improved performance metrics. Gemini 3.6 Flash, in particular, showcases a substantial enhancement by reducing output token usage by 17% compared to its predecessor, the 3.5 Flash. In some scenarios, such as the DeepSWE benchmark developed by Datacurve, the improvement in token efficiency reaches an impressive 65%.

The reduction in token usage is not the only achievement; Gemini 3.6 Flash also stands out in terms of precision and capability. In the DeepSWE benchmark, the model achieves 49% precision in code edits, surpassing the previous 37% attained by 3.5 Flash. Task performance is another area where the model excels, delivering a 63.9% improvement in ML research benchmarks versus the prior version’s 49.7%. Additionally, this model is more adept at tasks evaluated by OSWorld-Verified metrics, with a performance score of 83.0% compared to 78.4% from 3.5 Flash. Gemini 3.6 Flash shines in knowledge work scenarios as well, scoring 1421 in GDPval-AA v2, overshadowing the 1349 score of its predecessor.

Intertwined with the mainline Gemini 3.6 Flash, the new Gemini 3.5 Flash-Lite introduces speed and efficiency. Known as the fastest model in the 3.5 series, it delivers 350 output tokens per second according to the Artificial Analysis Index. Its performance improvements are evidenced by a 54% increase in Terminal-Bench 2.1 metrics over its previous iteration, and an impressive 72.2% boost in long context tasks per GDM-MRCR v2 results. Moreover, the 1140 score in real-world task execution benchmarks in GDPval-AA v2 further solidifies its capability with substantial gains from its preceding version’s 642 score.

Finally, the Gemini 3.5 Flash Cyber model is described as flying at the competitive frontier. This model represents a robust option for specific cybersecurity applications with integration through the CodeMender solution, combining state-of-the-art technology with Gemini’s latest innovations. As these models hit the market, they usher in an era where AI models are not only faster and more efficient but also tailored for precise applications, pushing the boundaries of AI capabilities.