Summary of Key Points
Recently, there has been a clear divergence in the pricing strategies of large model markets: on one hand, DeepSeek has increased its prices, while on the other hand, Google has launched the new Gemini 3.7 Flash model and offered a limited-time 50% discount (valid until the end of 2024). This new model boasts improvements in both intelligence score and speed, as well as lower costs. However, Google's flagship model, Gemini 3.5 Pro, has yet to be released. Additionally, there have been significant personnel changes at Google DeepMind, leading to speculation that the company may be shifting from a focus on developing the "top-of-the-line flagship models" to a more cost-effective and scalable approach. Although Google still holds core advantages such as its own developed chips and computing power, the uncertainty around the direction of its AI business has increased.
Detailed Analysis
1. Google's 50% Discount Offer: A Lucrative Deal for Developers
Google is offering a substantial discount on Gemini 3.7 Flash—by the end of this year, the cost per million tokens for input will be only $0.75, with an output cost of $3.75, which is half the price of the previous generation, 3.6 Flash. This discount is not permanent; the price will revert to its normal level in January 2027 ($1.5 per million tokens for input and $7.5 for output), effectively providing developers with a "five-months discount coupon." Developers are reacting enthusiastically, stating that this offer is incredibly affordable. For example, using this model to automatically restructure an entire Next.js codebase will not deplete their API budget, representing a significant victory for independent developers. Google's strategy is clear: by lowering the barriers to trial and adoption, it aims to make the Flash series the "default tool" in the era of intelligent models.
2. The Three Key Advantages of Gemini 3.7 Flash: Intelligence, Speed, and Cost Efficiency
The core strengths of this new model lie in its balanced approach to intelligence, speed, and cost:
- Intelligence: It scored 56 on tests (in high-complexity mode), which is 4 points higher than 3.6 Flash and surpasses domestic models like DeepSeek V4 Pro and GLM-5.2 (both scoring 53).
- Speed: The output speed is approximately 340 tokens per second, three times that of GPT-5.6 Terra and GLM-5.2; tasks can be completed in an average of just 1.7 minutes in high-complexity mode, making it the best model in terms of both intelligence and speed.
- Cost Efficiency: The cost per task in high-complexity mode is around $0.4, a 30% reduction from 3.6 Flash, while the cost in medium-complexity mode has been lowered to $0.26.
Google's positioning for the Flash series is clear: it does not aim to be the "smartest" model but rather the one most suitable for large-scale use.
3. The Delay of the Flagship Model: Is Google Changing Its Approach?
In May, Google announced that Gemini 3.5 Pro would be released the following month, but three months have passed without any news. In contrast, the Flash series has seen rapid updates from version 3.5 to 3.7 in just three months. There are rumors that Google may be adjusting its strategy, moving away from developing ultra-large flagship models with large parameter sizes and high costs towards a focus on the cost-effective Flash series, aiming to gain a competitive edge through increased usage and broader product coverage. Although Google has not confirmed these rumors, this shift makes sense: the Flash series is more suitable for general developers and large-scale enterprise deployments, allowing it to quickly capture market share. The higher development costs and slower rollout of flagship models may indicate a temporary shift towards practicality.
4. Major Personnel Changes at DeepMind: Uncertainty in AI Direction
Recent changes at Google DeepMind have raised questions about the company's AI strategy:
- Founder Demis Hassabis has relinquished day-to-day management responsibilities to become Chairman and Chief Scientist of Alphabet.
- Chris Carucaoğlu has been promoted to Senior Vice President, responsible for daily operations.
- Jeff Dean, a legendary engineer with 27 years at Google, has left to start his own company. There are also rumors that DeepMind may be undergoing reorganization.
Combined with the delay in the release of the flagship model, these developments have led to skepticism about Google's AI direction. While Google's previous successes with Gemini 3 Pro and Nano Banana were impressive, its current progress seems somewhat chaotic. However, these changes could also signal an attempt to adjust the company's AI business structure, and it will be important to watch the actions of the new team in the coming months.
5. Divergent Pricing Strategies in the Large Model Market
The current pricing strategies in the large model market are polarizing:
- Google's Discount Offer: Using lower prices to attract new users and expand its user base, thereby gaining a share of the intelligent model market.
- DeepSeek's Price Hikes: This could be due to cost pressures (high training and maintenance costs) or an attempt to target premium customers with higher purchasing power.
This divergence reflects the industry's ongoing exploration of profitable models. Some companies focus on scale, while others rely on premium pricing. Google's unique advantages, such as its custom chips (TPUs), ample computing power, and access to a vast user base through YouTube and search, give it a competitive edge that other firms struggle to match. The future landscape of the large model market remains uncertain.
In Conclusion
The large model market is shifting from a focus on "who is the smartest" to "who is the most practical." Google's discount strategy and the continuous iteration of the Flash series are clear signs of this transformation. However, whether Google will ultimately succeed depends on the long-term reliability of its models and their acceptance by the market.