Gemini 3.5 Flash + Pro: Powerful, Cheap, & Fast NEW AI Model! (Fully Tested)

Key Concepts

• Gemini 3.5 (Flash & Pro): The upcoming iteration of Google’s LLM series, internally codenamed "Cappuccino" for the Pro variant.
• Checkpointing: The process of testing specific model iterations (e.g., Fanta, Sprite, Cola) during development.
• Spatial Consistency: The ability of an AI to maintain structural integrity and relative positioning in complex outputs like ASCII art or code-generated interfaces.
• Token-level Planning: The model's capacity to map out complex, multi-part outputs (like game code) before execution.
• SaaS UI Over-training: A phenomenon where models default to repetitive, dashboard-heavy, or "modern" interface aesthetics regardless of user prompts.
• Prompt Adherence: The model's ability to strictly follow negative constraints (e.g., "do not use web access").

---

1. Evolution of Gemini Models

The video details the transition from the underwhelming Gemini 3.2 series to the highly anticipated Gemini 3.5 series.

• Gemini 3.2 Series: Initially tested via checkpoints like Fanta, Sprite, and Cola. The 3.2 Pro variant was criticized for lacking qualitative jumps and suffering from repetitive, "GPT-like" UI output styles.
• Gemini 3.5 Series: Set to launch at Google I/O. The new "Cappuccino" (Pro) and Flash variants demonstrate a significant leap in output quality, moving away from the repetitive panel-style UI of previous versions.

2. Technical Performance and Capabilities

The Gemini 3.5 Flash checkpoint has shown remarkable improvements in coding and structural generation:

• Minecraft Clone Generation: The model successfully generated a functional sandbox game clone, including background music, server connectivity, movement mechanics, and game modes (Creative vs. Survival). This represents a milestone in model capability regarding game dynamics and aesthetic structure.
• ASCII Art Generation: The model demonstrated high-level spatial consistency by generating a complex ASCII art piece of a pelican riding a bike. It allowed for user-defined presets, such as hue adjustments, character cell size, and teletype scan lines.
• Comparison: The 3.5 Flash output is described as competitive with, or superior to, Claude Opus 4.7 and Gemini 3.1 Pro in terms of layout hierarchy and instruction following.

3. Identified Weaknesses and Limitations

Despite the performance gains, the model exhibits specific technical flaws:

• UI Over-saturation: The model tends to force "useless" HUDs, floating panels, and dashboards into outputs, even when explicitly instructed to use a "minimalistic" design. This suggests the model may be over-trained on modern SaaS UI aesthetics.
• Prompt Adherence Issues: The model struggles with negative constraints, specifically regarding web access. When instructed not to use browsing tools, it often ignores the command, making it difficult to verify knowledge cutoffs.
• Inconsistency: While the 3.5 Flash is impressive, some complex tasks (like a Mac OS clone) showed mixed results compared to previous 3.2 Pro generations, suggesting that the model is still being fine-tuned.

4. Methodology for Testing

The testing was conducted via the "Alamarina" platform (likely referring to the LMSYS Chatbot Arena).

• Battle Mode: Users access these experimental checkpoints by entering "Battle Mode" in the arena, where they are randomly assigned either the new 3.5 Flash or older 3.2 checkpoints.
• Secret Testing: There are reports that some models labeled as "Gemini 3.1 Pro" in the arena are actually running the newer 3.5 Flash/Pro architecture, indicating that Google is conducting "blind" A/B testing to gather unbiased performance data.

5. Synthesis and Conclusion

The Gemini 3.5 series represents a significant qualitative shift for Google’s AI, particularly in its ability to handle complex, multi-component coding tasks and maintain spatial consistency. While the model shows "remarkable" improvements in reasoning and aesthetic hierarchy, it remains hampered by a tendency to over-engineer UI elements and occasional failures in strict prompt adherence. As Google I/O approaches, the primary takeaway is that the 3.5 Flash variant is positioning itself as a high-performance, cost-effective tool that significantly outperforms its predecessors in creative and structural generation.