Google has officially rolled out a new iteration of its AI model, Gemini 3.5 Flash (Low), aimed at optimizing token consumption on its agentic software development platform, Antigravity. The move comes as a direct response to growing user frustration over a newly structured, compute-based usage system that rapidly exhausted weekly token quotas—even during minor coding tasks.
Tackling the Antigravity Token Crisis
Following recent updates to the Antigravity IDE, users across both free and paid tiers reported hitting severe weekly lockouts within just a few hours or sessions. The platform’s advanced agentic workflows—which execute planning steps, run terminal commands, and inspect logs—were inadvertently consuming a massive volume of tokens for simple code edits.
In response, Google DeepMind introduced Gemini 3.5 Flash (Low) specifically tailored for low-effort, routine programming tasks. According to Varun Mohan, Director at Google DeepMind working on Antigravity, the new “Low” tier generates approximately 45% fewer tokens compared to the standard model, which has now been rebranded as Gemini 3.5 Flash (Medium).
Performance and Quota Adjustments
Despite its lower resource footprint, internal testing reveals that the “Low” model generally outperforms the older Gemini 3 Flash (now categorized as the “High” model) on Software Engineering (SWE) tasks. Under this new configuration, the Low, Medium, and High designations directly correspond to the depth of reasoning and compute power the model utilizes.
To give developers immediate breathing room as they adjust to the tiered framework, Google has simultaneously executed a full reset of Gemini quotas across all free and paid subscription plans. Moving forward, developers are encouraged to rely on the Low variant for standard code modifications to maximize their weekly limits, reserving the High and Medium variants exclusively for complex debugging and architectural planning
