Summary of Key Points
Google has just released three new Gemini models (3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber), claiming they are “more cost-effective, more efficient, and smarter.” However, both netizens and third-party evaluations have criticized them, stating that there is no improvement in intelligence; instead, there seems to be a regression in performance, and the cost-effectiveness is inferior to competitors’ models. The flagship model, 3.5 Pro, has even experienced delays in release. People jokingly refer to Gemini as if it has “Alzheimer’s disease,” becoming less intelligent with each update, which only leads to more laughter from users.
What Does Google Mean by “Cost-Effective and Efficient”?
Google describes these new models as being “mainstream solutions” that are “affordable.” The specifics include:
- Reduced Token Usage: The 3.6 Flash model uses 17% fewer tokens to produce the same amount of output compared to its predecessor, with a 65% reduction in complexity for certain tasks. In simple terms, it generates less unnecessary text when answering questions, meaning users spend less money for the same results. For example, an answer that used to require 100 tokens now only requires 83.
- Price Reduction: The price of the 3.6 Flash model has been lowered from $9 per million tokens to $7.5, while the Flash-Lite model is even more affordable ( costing $0.3 for input and $2.5 for output). These models are designed for high-throughput, low-latency tasks such as batch document processing and searching.
- Minor Improvements: The knowledge update date has been extended from January 2025 to March 2026, and there have been slight improvements in code capabilities and computer interaction (such as automatic software usage), along with enhanced security features against chemical and cyber attacks.
However, these claims are met with skepticism from users who have not seen a significant improvement in actual performance.
Why Are Users Laughing So Hard?
Third-party evaluations and user tests have thoroughly debunked Google’s claims:
- No Improvement in Intelligence: Independent organizations have found that the intelligence level of the 3.6 Flash model is the same as its predecessor, or even inferior to competitors like Meta’s and GLM’s models. Users complain that it is more expensive and less intelligent than GPT models, which can perform better for the same tasks at a lower cost.
- Numerous Issues in Tests: Users have reported various problems, including basic decoding errors, poor Chinese language usage, low-quality generated images and videos that do not match instructions, and frequent confusion or misuse of tools. Some even suspect that Google may have sold its computing resources to others, leading to these shortcomings.
- Flash-Lite Outperforms the Older Model: Interestingly, the lower-tier Flash-Lite model has surpassed the older 3 Flash model in some tasks, highlighting the weakness of the older versions.
The Delay in the Release of the Flagship Model (3.5 Pro)
Where has the highly anticipated 3.5 Pro model gone? Google claims it is still in testing, but the truth is less impressive:
- Internal Failures: According to reports, the code generation capabilities of 3.5 Pro have not met Google’s own standards. After updating the training data in June, issues persisted, necessitating a complete restart from scratch.
- Diversion of Attention: Instead of focusing on 3.5 Pro, Google is promoting Gemini 4, claiming it involves “the most ambitious pre-training effort to date.” This move seems more like a attempt to shift attention away from the shortcomings of 3.5 Pro.
Poor Cost-Effectiveness: The Foundation of the Flash Series Is Shaken
The Flash series was once known for its cost-effectiveness, but this has been severely questioned:
- More Expensive and Less Powerful: Users have compared Gemini 3.6 Flash to GPT 5.6 Sol medium and found that it is more expensive while offering less intelligence. For example, GPT can complete similar tasks more efficiently at a lower cost, eroding the Flash series’ competitive advantage.
- User Disapproval: Users are disappointed, stating that the combination of high price and poor performance is unacceptable. In contrast, OpenAI’s Codex and ChatGPT continue to see increasing usage, highlighting Google’s falling position in the market.
The Secure Model (3.5 Flash Cyber) Is Highly Restricted
The latest model, 3.5 Flash Cyber, is designed to detect code vulnerabilities but is strictly controlled by Google:
- Limited Access: This model is only available for “trusted partners” to prevent it from being used for malicious purposes, such as hacking attacks. Google’s approach assumes that AI can identify vulnerabilities faster than they can be fixed, so it aims to provide a defense before attackers can exploit them.
- The Dilemma of Technology: While this is a useful feature, it also highlights the challenges associated with AI security, as misuse could have serious consequences.
Conclusion: Will Google’s Jokes About “Alzheimer’s Disease” Come True?
Google’s new models appear to be more expensive and less intelligent than before, with the flagship model facing delays in release. The jokes about Gemini becoming less intelligent are not unfounded. From last year to now, there has indeed been a noticeable decline in its performance. If 3.5 Pro fails to meet expectations, the “Alzheimer’s disease” analogy may become a reality. Google’s current strategy seems more like a effort to maintain stability, but whether it can regain user trust remains to be seen.
(End of translation)