谷歌发布三款Gemini新模型 长任务Token成本最高降65%

Google DeepMind has launched three new proprietary Gemini AI models, with core goals of improving the speed and intelligence of large-scale AI agent operations while drastically cutting usage costs. Key takeaways are as follows:

1. The three models are positioned to serve different use cases, with sharp price cuts compared to the previous generation. Token costs for long tasks are reduced by up to 65%, and performance is improved across the board. Both general-purpose models come with a 1 million-token context window and an enhanced security mechanism that lowers abuse risks while reducing false blocks on legitimate use cases.

2. The two general models are currently accessible via multiple official Google channels. The cybersecurity-specific model Gemini 3.5 Flash Cyber will only be available to governments and trusted partners moving forward. All models are closed-source and only accessible via API; enterprises that require on-premises deployment must sign a Google Cloud advanced enterprise agreement.

3. Google has outlined its clear product roadmap. The previously announced flagship model Gemini 3.5 Pro is already in testing with partners, and pre-training for the next-generation flagship Gemini 4 has begun, marking the continued acceleration of large model product iteration cycles.

Google has released three new Gemini models, opening up new opportunities for digital operations and product innovation for brands of all types. Key key takeaways for brands are as follows:

1. Lower technology costs have reduced the barrier to implementing AI applications for brands. Token costs for long tasks on the new models drop by up to 65%, and the lightweight model runs twice as fast as its predecessor. Brands can now integrate AI at a much lower cost for content creation, user data analysis, market research, customer service and other workflows, effectively cutting operating costs.

2. Improved performance of the new models allows brands to handle complex enterprise-level requirements, such as parsing complex documents, analyzing chart data, and writing long reports. It also supports advanced use cases like complex code migration and 3D tool development. Brands can leverage AI capabilities to optimize internal workflows and launch smarter products and services that better meet user needs.

3. The iteration pace of the large model industry has clearly accelerated. Competitors have already updated their flagship products multiple times, and Google is also speeding up its product release cycle. Rapid advancements in AI technology will continue to drive new consumer trends, so brands need to adjust their product and operation strategies to keep up with technological changes.

Google's release of three new Gemini models brings new opportunities and implications for both AI-focused sellers and traditional sellers that use AI for operations. Key takeaways for sellers are as follows:

1. On the opportunity side: The new models deliver major cost reductions and noticeable speed improvements, allowing small and medium-sized sellers to access AI capabilities at lower costs for daily operations including product selection analysis, customer service responses, content creation, and product listing writing. This lowers the barrier to AI adoption, enabling large-scale deployment of long-cycle complex AI applications that were previously out of reach due to cost constraints.

2. Access rules are clear: Regular sellers can directly access the models via official channels such as Google AI Studio on a pay-as-you-go basis, with no high upfront investment required and a very low entry barrier. Enterprise sellers with special requirements for on-premises deployment can achieve this by signing an advanced agreement with Google Cloud.

3. Risk note: The large model industry is currently iterating extremely rapidly. Competitors have updated their flagship products multiple times, and Google's flagship model is still in testing. With fast technological updates, sellers need to keep up with iterations and upgrade the AI tools they use in a timely manner, while strictly complying with Google's usage policies and rate limits.

Google's release of three new cost-effective Gemini models brings new insights and opportunities for factories advancing digital and intelligent transformation. Key takeaways for factories are as follows:

1. For product production and design: The new large models deliver lower costs and improved performance, enabling factories to use AI for product design optimization, engineering drawing analysis, and production process improvement. Complex design-focused AI applications that previously required high costs can now be deployed at a much lower price point, helping factories shorten R&D and design cycles and cut design costs.

2. Insights for digital transformation: Lower costs and improved performance have drastically lowered the barrier for factories to deploy AI at scale. Factories can use the new models to handle high-load work including production data parsing, quality inspection report analysis, and supply chain process optimization, effectively improving production and operating efficiency.

3. New business opportunities: With faster large model iteration, many AI applications for the manufacturing sector can be developed rapidly based on the new models. Factories can either integrate the new model capabilities to build intelligent tools tailored to their own needs, or partner with AI service providers to develop vertical AI applications for manufacturing, opening up new business lines.

Google's release of three new Gemini models provides a lot of valuable industry insights for all types of AI service providers. Key takeaways for service providers are as follows:

1. Industry development trends: The large model industry is currently iterating at a clearly accelerated pace, with leading players releasing product updates at high frequency. Industry competition is focused on reducing usage costs, improving operating speed, and optimizing efficiency for long tasks. At the same time, specialized models for vertical and niche segments have emerged as an important new development direction, with the new cybersecurity-specific model launched this time being a typical example.

2. Core customer pain points: Customers' demands for cost control, high concurrency and low latency for large models are growing increasingly prominent. Previous generation models were slow and expensive, and could not meet customers' demand for running large-scale AI agents. In addition, customers have rising requirements for large model security: they need both to reduce abuse risks and cut down on false blocking of legitimate use cases.

3. Solution directions: Service providers can build tailored solutions for high-throughput low-latency scenarios, high-load complex task scenarios, and vertical specialized fields respectively, leveraging the tiered capabilities of Google's new models to meet customers' differentiated needs. Meanwhile, the new models are accessible via open API, allowing service providers to integrate capabilities quickly without training large models in-house, which drastically cuts development costs.

Google's release of the new Gemini models provides a lot of valuable operational insights for large model platforms and AI application platforms. Key takeaways for platform operators are as follows:

1. Market demand trends: Developers and enterprise clients are highly sensitive to large model costs, and cost reduction is currently the core demand. They also have clear differentiated requirements for operating speed, long-task processing capabilities, and vertical scenario adaptation. Platforms need to add models of different positioning to cover all types of needs, from cost-effective lightweight scenarios to high-load complex scenarios.

2. Operational management reference: Google has implemented tiered access rules for models of different natures: general-purpose models are fully open to meet the needs of regular customers, while specialized models for sensitive fields are only available to trusted partners to maintain security. This tiered open access model is well worth adopting for other platforms.

3. Guidance on risk mitigation: Large model security is a core regulatory requirement. The new Gemini models strengthen jailbreak protection to reduce abuse risks in sensitive fields, while optimizing blocking rules for legitimate use to balance security and user experience. Platforms can refer to this approach to optimize their products. They also need to keep up with the industry's iteration pace and add new models in a timely manner to avoid falling behind competitors.

Google's release of three new Gemini models reveals a number of new trends in the large model industry, offering high reference value for industry researchers. Key insights are as follows:

1. New industry trends: The direction of large model product iteration is now clearly diverging. On one hand, general-purpose models continue to cut costs and improve efficiency through technical optimization, reducing token consumption to lower long-task costs and boost operating speed, to meet the development needs of large-scale AI agents. On the other hand, players are investing in specialized models for vertical niche segments, fine-tuning models for specific fields to fit the professional needs of niche use cases, with the new cybersecurity-specific model being a typical example.

2. New changes in industry competition: Competition among leading large model vendors is intensifying. OpenAI and Anthropic have already updated their flagship products multiple times, outperforming Google's current flagship in performance. Google has therefore accelerated its product iteration pace: it is already conducting partner testing of the new flagship Gemini 3.5 Pro, and has started pre-training for the next-generation Gemini 4, leading to a clear overall increase in the industry's product iteration speed.

3. In terms of business model, the current mainstream approach for leading vendors remains closed-source pay-as-you-go API access, with on-premises deployment only available to high-tier enterprise clients that require it, and no model weights or source code released. This model protects vendors' revenue while meeting the differentiated needs of different customers, and has become the dominant business model for proprietary large models today.