Google has released a lightweight variant of its Gemini 3.5 Flash model designed to solve a growing frustration among developers: token consumption that drains usage quotas within minutes during intensive coding tasks. The new Gemini 3.5 Flash (Low) generates approximately 45% fewer tokens than the standard Gemini 3.5 Flash (Medium) version, offering a practical solution to the token shortage that emerged after Google shifted to a compute-based pricing model.

What Caused Google's Token Shortage Problem?

The trouble began when Google transitioned its Gemini pricing from a daily prompt-based system to a compute-based usage model. Under the new system, the platform measures consumption based on which AI features users employ, the complexity of individual requests, and the length of chat history. For developers using Antigravity, Google's agentic coding platform, this shift meant that complex coding tasks consumed enormous amounts of computational resources, exhausting weekly token quotas in minutes.

The pricing change sparked significant backlash from the developer community, forcing Google to expand usage limits for Gemini Pro models and rethink its approach. Rather than simply raising quotas across the board, Google opted to create a tiered variant system for Gemini 3.5 Flash, allowing developers to choose the right model for their specific task complexity.

How Does Google's New Tiered Gemini Approach Work?

Google is now developing three variants of Gemini 3.5 Flash, each optimized for different use cases. The original model is now called Gemini 3.5 Flash (Medium), and it sits in the middle of a new spectrum. Here's how the tiers break down:

• Gemini 3.5 Flash (Low): Optimized for simple tasks with minimal token overhead, generating 45% fewer tokens than the Medium version and outperforming the earlier Gemini 3 Flash (High) on software engineering work.
• Gemini 3.5 Flash (Medium): The original Gemini 3.5 Flash model, suitable for standard coding and development tasks with balanced performance and token efficiency.
• Gemini 3.5 Flash (High): Designed for complex tasks requiring more computational resources and deeper reasoning capabilities.

This tiered approach reflects a broader industry trend toward offering multiple model sizes for different workloads. Gemini 3.5 models are four times faster than frontier models and significantly less costly, making them attractive for developers who need speed without premium pricing.

"We heard concerns that Antigravity consumes many tokens for simple tasks now. So, we're adding Gemini 3.5 Flash (Low) as a way to optimize token usage for these tasks. In our internal testing, it generates around 45% fewer tokens than Gemini 3.5 Flash (Medium) and generally outperforms Gemini 3 Flash (High) on software engineering tasks," stated Varun Mohan, Google DeepMind Director.

Varun Mohan, Google DeepMind Director

What Other Changes Did Google Make to Address Developer Concerns?

Beyond releasing the Low variant, Google announced updated token quotas across both free and paid Gemini plans. These adjustments aim to ensure developers can continue their software engineering work without hitting usage walls unexpectedly. However, developers have raised additional concerns about rate limits on Google's image generation capabilities within Antigravity.

Some users pointed out significant disparities between Antigravity's image generation limits and competing platforms. One developer noted being able to generate 1,000 images on a competing service while only generating 24 on Antigravity's Ultra plan. When asked about these limits, Mohan acknowledged they were "pretty low" and that it "makes sense to increase" them, though he did not commit to a specific timeline for raising those caps.

Why This Matters for Enterprise AI Adoption

The token shortage and Google's response highlight a critical challenge in enterprise AI adoption: matching model capabilities to real-world usage patterns. Large language models (LLMs) are general-purpose systems trained on vast text datasets that can handle multiple tasks including text generation, summarization, question answering, code generation, and information extraction. However, deploying these models at scale requires careful attention to resource consumption and cost management.

For organizations building on Google's platform, the availability of multiple Gemini 3.5 Flash variants means they can now optimize their infrastructure spending by matching model size to task complexity. A simple code completion task no longer requires the same computational overhead as a complex architectural review, potentially reducing costs significantly for high-volume use cases.

The broader lesson extends beyond Google: as AI models become embedded in production workflows, developers and enterprises need flexibility in model selection. The shift toward tiered offerings reflects industry recognition that one-size-fits-all models don't align with real-world deployment economics.