Google has launched three new AI models — Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber — but the model the market was waiting for, Gemini 3.5 Pro, still lacks a release date. The company attributes this to it not passing internal tests, particularly in code generation, as it confirmed to the Chinese publication 36Kr. The strategy is clear: saturate the market with niche variants to disguise the fact that the flagship product isn’t ready. But the most revealing piece of data isn’t in the models—it’s in the silicon: a leak of the internal “Frozen v2” chip, which promises to multiply the efficiency of current TPUs, suggests Google isn’t simply delaying a launch but redesigning its entire architecture from the ground up. The question for anyone investing, competing, or regulating this sector is whether we’re witnessing a tactical stumble or the symptom of a structural crisis at Mountain View’s lab.
A Flood of Models That Doesn’t Mask the Central Void
TechCrunch put it plainly: “Google releases three new Gemini models — but no 3.5 Pro.” The absence of the most powerful model, which according to the internal roadmap should have arrived earlier, has become the elephant in the room in every conversation about artificial intelligence. The three models released—Gemini 3.6 Flash, geared toward fast responses and low computational cost; Gemini 3.5 Flash-Lite, an even lighter version for devices with limited resources; and Gemini 3.5 Flash Cyber, specialized in cybersecurity—are variants of an ecosystem branching out while the main trunk remains stagnant.
Google told 36Kr that Gemini 3.5 Pro is being tested with partners and will launch “soon,” without providing a specific date. The information, also reported by the Chinese publication Interface (界面), indicates that the model failed to meet its goals in internal tests, especially in code generation. This detail is significant: the ability to generate functional, complex code has become the battlefield where the true power of frontier models is measured. Meanwhile, OpenAI has already launched GPT-5.6 Sol, and the Chinese lab Moonshot AI has introduced Kimi K3, both with standout performance in programming benchmarks.
The Problem Isn’t Timing—It’s Architecture
That a model is delayed is not, in itself, extraordinary news. What’s concerning for Google is that the delay of Gemini 3.5 Pro is not an isolated incident but the third significant postponement in less than a year. Earlier, the company had already delayed the launch of Gemini 3.0 Pro due to similar issues. And later, the intermediate version Gemini 3.2 Pro also suffered delays.
The pattern suggests the problem isn’t timing but architecture. While Flash models—distilled, optimized versions for specific tasks—can be iterated relatively quickly, the Pro model requires an internal coherence that seems to be eluding Google’s teams. Code generation, in particular, demands a deep understanding of logical structures, syntax, and context that Google’s current models cannot stabilize. According to engineers cited by 36Kr, Google’s teams have had to rewrite significant parts of Gemini 3.5 Pro’s training pipeline, delaying the entire timeline.
In this context, the decision to launch three niche models has a strategic reading: Google needs to show its ecosystem is still alive, that there are updates, that the lab hasn’t stopped. But for analysts closely tracking the sector, the signal is the opposite: the more minor models are released, the more glaring the absence of the big one becomes.
Frozen v2: The Silent Bet That Changes Everything
If the launch of minor models is the smokescreen, the leak of the “Frozen v2” chip is the real news. According to 36Kr, Google is developing an internal chip that permanently embeds part of Gemini’s architecture into silicon, reducing the computation and data transfer needed during inference. Google engineers estimate this design offers token processing capacity per unit of power several times higher compared to current TPUs.
The impact on markets was immediate: Google’s shares (GOOGL and GOOG) rose after the leak was reported. Investors interpreted that Google isn’t just patching its model but redesigning the entire tech stack, from silicon to the application layer. If Frozen v2 delivers on its promise, Google could drastically reduce the inference cost of its largest models, giving it a competitive advantage in price and scalability that neither OpenAI nor Chinese labs can match in the short term.
However, this bet has an obvious risk: developing custom chips is slow, expensive, and not always successful. Google has already faced issues with previous TPU generations, and Frozen v2 appears to be a radical redesign, not an incremental evolution. While the chip isn’t ready—and there’s no mass production date—Google is competing with one hand tied behind its back.
The Pulse of the Frontier: Who Sets the Pace
Google’s situation can’t be understood without the global context. OpenAI launched GPT-5.6 Sol, with code performance that far surpassed Gemini 3.0 Pro. In China, Moonshot AI presented Kimi K3, with multimodal capabilities that have surprised even the most skeptical analysts. And Anthropic, though more low-key, has steadily improved its Claude 4 model.
Google, which for years was the leading lab in artificial intelligence, has moved to the defensive. Its strategy of fragmenting the market with specialized models might work in the short term to retain enterprise clients needing concrete solutions, but it doesn’t resolve the fundamental question: Can Google produce a frontier model that competes head-to-head with GPT-5.6 Sol or Kimi K3?
The answer, today, is uncertain. And the silence surrounding Gemini 3.5 Pro, which was supposed to be that answer, only fuels the doubts.
A Look Ahead: The Risk of Playing Defense
Google is not a small company, nor is it in immediate danger. Its advertising revenue, cloud business, and Android ecosystem give it a financial cushion that no other AI lab has. But in the race for the AI frontier, money isn’t everything. The ability to attract talent, maintain research team morale, and execute quickly is what separates leaders from followers.
Google’s current strategy—stocking the storefront with minor models while redesigning its silicon—could be brilliant if Frozen v2 arrives on time and Gemini 3.5 Pro finally works. But it could also be a suicidal gamble if the chip is delayed or fails to meet expectations, and the flagship model still fails to arrive while OpenAI and Chinese labs push ahead.
For our global audience, the lesson is clear: in the geopolitics of artificial intelligence, delays are not accidents—they are signals. Google is losing the pulse of the frontier, and its response isn’t to run faster but to swap horses. We’ll see if the new horse—Frozen v2—can gallop before the race ends without it.