The video discusses the Gemini Flash 3 model, highlighting it as the creator’s new favorite AI model despite initial skepticism about its predecessor, Gemini 3 Pro. The speaker has been using Google’s 2.5 Flash model daily for nearly a year and appreciates its balance of speed and capability, despite some quirks. Gemini 3 Flash represents a significant performance upgrade over 2.5 Flash, outperforming many other models on benchmarks like Skatebench, especially in spatial reasoning, image, video, and audio processing. The model is praised for its efficiency, speed, and advanced multimodal capabilities, making it suitable for complex tasks such as game development, deepfake detection, and document analysis.
The video also covers the pricing and cost-efficiency of Gemini 3 Flash. While it is more expensive than previous Flash versions, with input and output token costs significantly increased, it remains much cheaper than the 3 Pro model. The model’s ability to reason extensively results in high token usage, which drives up costs but also contributes to its intelligence and performance. Despite this, Gemini 3 Flash offers a compelling price-to-performance ratio, especially for bulk data processing tasks where its batch API and high rate limits provide cost savings and efficiency.
A notable downside of Gemini 3 Flash is its high hallucination rate, with the model reportedly fabricating answers 91% of the time when it does not know the correct response. This tendency to “lie” rather than admit ignorance poses challenges for applications requiring high accuracy and reliability. The speaker emphasizes the importance of understanding this limitation and building appropriate safeguards when integrating the model into workflows. The model also exhibits a strong inclination to “snitch” or report undesirable behavior when tested with scenarios involving ethical dilemmas, which reflects its programming but may affect its use in sensitive contexts.
The speaker compares Gemini 3 Flash to other models in terms of instruction-following and usability. While Gemini models excel in knowledge and reasoning, they lag behind competitors like OpenAI and Anthropic in following instructions precisely and avoiding unnecessary deviations. The model tends to overcomplicate tasks, such as generating excessive code or UI elements beyond what was requested. Despite these quirks, Gemini 3 Flash is highly valued for specialized use cases involving data parsing, spatial reasoning, and multimodal inputs rather than casual chat or coding assistance.
Finally, the video criticizes the user experience around Google’s AI platforms, describing AI Studio and Vertex as difficult to use and poorly designed, which detracts from the otherwise impressive capabilities of Gemini models. The speaker recommends using third-party tools like Open Router or T3 Chat to access Gemini models more effectively. Overall, Gemini 3 Flash is portrayed as a powerful, cost-effective model best suited for complex data processing tasks rather than everyday conversational or coding applications. The speaker expresses enthusiasm for the model’s potential and encourages viewers to experiment with it while acknowledging its current limitations.