Gemini 3.5 Pro Delayed Again: Google's Compute Dilemma and the Mystery of a Full Rebuild

Another Delay: The Community's Mixed Feelings

Recently, news of "Gemini 3.5 Pro being delayed again" circulated within the Reddit community, sparking heated discussion among AI enthusiasts. What you might not have noticed is that the overall community sentiment wasn't anger, but rather a kind of tolerance tinged with anticipation. Several users expressed: "As long as this model is genuinely stronger than the current flagship, waiting a few more weeks is fine."

This attitude reflects a subtle shift in the current competitive landscape of large AI models. Google DeepMind's long-term investment in multimodality, world knowledge, and embodied world modeling has given many users confidence in the company, even though the Gemini series doesn't currently sit firmly at the cutting edge of capability.

What is embodied world modeling? Embodied world modeling refers to an AI system's ability to build an internal representation of a real environment by simulating perception-action loops in the physical world. Unlike pure text or image understanding, embodied AI needs to understand spatial relationships, physical laws, and causal chains, thereby supporting downstream tasks such as robotic manipulation and autonomous driving. Representative work from Google DeepMind in this direction includes Gato (2022, which for the first time unified 672 different tasks within a single Transformer network) and RT-2 (2023, which achieved end-to-end integration of vision-language models with robotic control policies), among other multi-task embodied agents. These systems can unify vision-language understanding with robotic arm control within the same neural network, without needing to design a separate controller for each task, and are considered one of the important paths toward artificial general intelligence.

Worth noting is that the core challenge of embodied world modeling lies in enabling AI to establish a robust causal model—not only describing the state of the world, but also predicting "if I execute action A, how will the world change." This requires breakthroughs across multiple subfields including visual perception, physical simulation, and natural language instruction understanding, making it far more complex than pure language or image understanding. Precisely because of this, research progress in this direction typically requires years of accumulation before it can be translated into product-level capabilities—but once a breakthrough occurs, it creates a competitive barrier that other approaches struggle to replicate.

It should be noted that model codenames mentioned in the discussion, such as "Fable" and "GPT-5.6-Sol," are mostly unofficial names or speculation circulating within the community. Some information was even directly called out by commenters as "completely fabricated"—for instance, the so-called "2M context length" has yet to be confirmed by any authoritative channel. Therefore, this article is more an observation of community sentiment and industry trends than a confirmation of specific product specifications.

From "AI Innovation Hub" to "Compute Server"?

The most noteworthy viewpoint in the discussion was one user's concern about Google's shifting role:

"Google is slowly but surely transitioning from an AI innovation hub into a compute server."

This statement pinpoints the core contradiction faced by all frontier labs: the tradeoff between capability and cost.

The Real Pressure of Compute Consumption

According to feedback based on community users' actual experiences, those truly impressive top-tier models often come with "horrifying" compute consumption. One comment mentioned that a certain competitor is "desperately trying to suppress usage of its high-end model"—controlling costs through measures like counting it as double the quota and restricting access to API calls only.

This reveals a harsh truth: the experience of top-tier models is essentially built on massive amounts of compute. For free users, this means stricter usage limits.

The real logic of compute economics: The cost of training and running inference on top-tier large models rises exponentially. Taking a GPT-4-class model as an example, the computational cost of a single inference is tens of times that of GPT-3.5, which directly puts providers under enormous marginal cost pressure. To balance commercial sustainability, platforms generally adopt a "compute tiering" strategy: restricting the strongest models to paid or API channels while offering reduced-capability versions to free users. The free Gemini 1.5 Pro calls previously offered by Google AI Studio were viewed by the industry as a rare generous strategy, but as inference costs continue to climb, the sustainability of this strategy is being questioned by the market.

From a deeper structural perspective, the superlinear growth in inference costs for large models based on the Transformer architecture has a clear mathematical root: the computational complexity of the self-attention mechanism grows quadratically with sequence length (O(n²)). Since the Transformer architecture was proposed by the Google Brain team in the 2017 paper Attention Is All You Need, its multi-head self-attention mechanism, while achieving powerful capture of global sequence dependencies, has also baked in this cost structure—meaning that when the context window expands from 128K to 1M, attention computation theoretically increases by about 60x. Although engineering optimizations like FlashAttention and sparse attention can partially alleviate this, the real cost increase still far exceeds linear expectations. For platform operators, offering long-context flagship models to free users is essentially a form of large-scale subsidy, whose sustainability is heavily dependent on venture capital funding and cross-subsidization from paying users. Once capital markets tighten their tolerance for "burning money to acquire users" strategies, the contraction of free quotas is almost inevitable. Notably, it was the Scaling Laws systematically quantified by OpenAI in 2020—where model capability grows in a predictable power-law relationship with parameter count, data volume, and compute—that gave rise to the entire industry's consensus on "brute-force parameter scaling," and fundamentally locked in the basic trajectory of this compute arms race.

Some users worry that as Google needs to devote more compute to high-end services, the free flagship call quotas currently enjoyed in AI Studio may be further compressed.

Will Google "Burn Compute" to Win?

If Google is willing to burn compute to win the AI race, it would undoubtedly be a boon for consumers. But most commenters take a cautious stance, generally expecting the new model's capability to fall "somewhere between mid and high-end, leaning slightly toward the more mature and stable end."

A Full Rebuild: Opportunity or Risk?

A major rumor circulating within the community is: Gemini 3.5 Pro may be a complete rebuild (full rebuild), rather than an enhanced iteration based on 2.5 Pro like 3.0 and 3.1.

What a Rebuild Means

If the rumor is true, this would be a double-edged sword. As one user analyzed:

"If it's a full rebuild, then it's an unknown quantity. It might be very impressive in capability, but also more prone to various quirks, bugs, and strange behaviors... great when it works, but potentially unreliable."

This "high ceiling, low stability" characteristic is the typical cost of any architecture-level update.

The engineering logic of rebuild vs. iteration: In large model development, a "full rebuild" typically means a comprehensive redesign spanning pretraining data mixtures, network architecture, and alignment strategies, while "incremental iteration" involves fine-tuning or local structural optimization on top of an existing checkpoint. The history of large model development offers ample evidence that architecture-level rebuilds coexist with capability leaps: OpenAI's transition from GPT-3 to GPT-4 involved key architectural explorations including Mixture-of-Experts, while Anthropic's move from Claude 2 to Claude 3 deepened its comprehensive Constitutional AI alignment strategy—the core idea of Constitutional AI being to preset an explicit set of principles for the model, allowing it to internalize compliant behavior through self-critique and revision, thereby reducing reliance on massive amounts of human-annotated feedback, forming a competitive alternative to the RLHF approach championed by OpenAI. Both experienced a phase of "capability leaps but initial instability," partly because new alignment recipes profoundly change how a model behaves in edge cases, giving it a completely different "personality" from previous versions. The core engineering challenge of a rebuild lies in the fact that the behavior of a new architecture under Scaling Laws is difficult to precisely predict from small-scale experiments, causing flagship models to often exhibit unexpected failure modes in specific specialized scenarios during their initial launch. This is also the reasonable logic behind Google choosing a prudent delay rather than a hasty release.

More specifically, a full architectural rebuild involves not only changes to the network structure, but also: re-screening and cleaning of the pretraining corpus (which determines the model's knowledge boundaries), comprehensive adjustment of alignment strategies (which affects the model's safety and instruction-following capability), and the redesign of post-training processes such as RLHF/RLAIF (which shapes the model's interaction style and refusal boundaries). The entire cycle typically takes several months to a year, and each stage may introduce new unexpected behaviors. This also explains why the industry commonly observes that "during the initial launch of a rebuilt version, professional users' experience is often worse than that of non-professional users"—because the former are more likely to hit failure zones in edge cases.

Some pragmatic users also expressed that even if 3.5 Pro is just "3.1 Pro plus better instruction following, extensive underlying bug fixes, and capability improvements in scenarios like agent tools," that would be perfectly acceptable.

Stable Iteration vs. Aggressive Innovation

This reflects two typical demands among the user base: some hope to see stable, reliable, and smoothly-experienced product-level improvements; others support Google going "all in" to create the truly strongest model. The tradeoff between these two approaches is precisely the strategic choice that every AI lab must face.

The Competitive Landscape: Why We Need Google

A rather profound viewpoint in the discussion concerns reflection on the necessity of competition.

One user bluntly stated: "Having OpenAI and Anthropic as the only labs standing at the frontier is not reassuring. They need competitors."

From a game-theoretic perspective, this judgment has a solid theoretical basis. The core insight of the Nash Equilibrium is that in a multiplayer game, when each participant has chosen their optimal strategy and cannot benefit by unilaterally changing it, the system reaches a stable state—and this equilibrium brings the most significant improvement in user welfare when the number of competitors increases from monopoly to two, with diminishing improvement but increasing system stability at three or more. The current tripartite competitive landscape forms a dynamic Nash Equilibrium: a capability breakthrough by any one party forces the other two to accelerate their R&D pace, thereby generating positive competitive externalities. Historically, single dominant institutions like Bell Labs and Xerox PARC once produced dense innovation in the short term, but over the long run often fell into path dependence due to the lack of external pressure—this provides an important historical reference for competitive diversification in the AI field.

DeepMind's Long-Termism

Supporters of this viewpoint further elaborated on why they are bullish on Google DeepMind:

• A firm commitment to omnimodality: not limited to text, but covering images, audio, video, and other modalities
• Deep accumulation of world knowledge: built on Google's vast data ecosystem
• Embodied world modeling: this direction is believed to bring significant returns over the long term

The technical background of multimodal competition: A multimodal large model refers to a unified model architecture capable of simultaneously processing text, images, audio, video, and other data types. Compared to the earlier approach of modeling each modality independently, unified multimodal models have significant advantages in cross-modal reasoning and information fusion. In the current competitive landscape, OpenAI's GPT-4o, Google's Gemini Ultra, and Anthropic's Claude 3 series are all evolving toward omnimodality, but each has different emphases in data scale, architecture design, and inference efficiency. Google, benefiting from massive multimodal data assets like YouTube and Google Search, possesses a data barrier in this arena that other labs struggle to replicate.

The value of this data barrier is often underestimated in long-term competition, but its magnitude is impressive: YouTube adds over 500 hours of new video content every minute, with a cumulative storage of over 1 billion hours of video-language paired data; Google Image Search indexes over a trillion images and their contextual semantic information; while Google Scholar and Google Books cover a structured knowledge graph of hundreds of millions of academic documents. The key value of these data assets lies not only in their scale, but also in the temporal causal annotations they naturally carry—video data implicitly encodes the laws of object motion, human behavioral logic, and causal chains of the physical world, making them native material for training embodied world models that cannot be replaced by ordinary web crawling. In comparison, the inherent deficiencies of OpenAI and Anthropic in accumulating native multimodal data may become an implicit constraint on their capability ceilings in the coming years. It's worth mentioning that this data moat has a deep coupling effect with Scaling Laws: under conditions of comparable compute and parameter count, data quality and diversity often become the key variable determining a model's capability ceiling, and Google's advantage in native multimodal data is precisely where it can be fully leveraged along this dimension.

Commenters called DeepMind "the most pragmatic and mature of the three major Western AI developers," and stated that although Gemini currently "still lags behind the frontier," as a long-term Gemma user, they firmly believe the Gemini 3.5 and even 4.0 series will grow into "truly powerful models."

Conclusion: Anticipation and Rationality Coexist

From this community discussion, we can see an increasingly mature mindset among today's AI users:

They are both full of anticipation for new models and soberly aware of real-world constraints such as compute costs, architectural risks, and free quota compression; they are both loyal users of a certain product and rationally acknowledge the strengths of competitors, deeply understanding that multi-party competition is the guarantee of healthy industry development.

Regarding the delay of Gemini 3.5 Pro, rather than waiting anxiously, it's better to view it as a signal of Google prudently weighing between "stable iteration" and "aggressive rebuilding." A truly powerful and reliable model is worth waiting a few extra weeks for.

Special note: The specific model codenames, parameters (such as context length), and release times mentioned in this article all come from community rumors and have not yet been officially confirmed. Readers are advised to reference them rationally.

Key Takeaways

• Maturing community sentiment: Users show anticipation-tinged tolerance toward the delay, reflecting subtle shifts in the AI competitive landscape
• Compute economics constraints: The O(n²) complexity of the Transformer attention mechanism leads to superlinear growth in inference costs. This architectural gene, laid down by the 2017 Attention Is All You Need, is structurally forcing platforms to contract free quotas—a trend dictated by mathematical law rather than a short-term strategy
• The double-edged effect of rebuilding: A new architecture may bring capability leaps, but initial stability risks are a historical pattern (see the evolution paths of GPT-3→4, Claude 2→3). The comprehensive reset of alignment strategies is an especially important source of uncertainty, and users need to adjust their expectations
• The long-term value of data barriers: Google's inherent advantage in native multimodal data (YouTube's 1B+ hours of video, trillion-scale image-text indexes) has a coupling effect with Scaling Laws, and may gradually manifest as structural differences in capability ceilings in the coming years of competition
• The necessity of competitive diversification: From a game-theoretic perspective, the dynamic Nash Equilibrium of a tripartite standoff is superior to a monopoly for both technological progress and user welfare. Historical cases like Bell Labs also confirm the risk of single dominant institutions falling into long-term path dependence; a healthy AI ecosystem requires multiple frontier labs advancing in parallel