Google's latest attempt to fix token quotas is here: Say hello to Gemini 3.5 Flash Low

Google’s latest Gemini 3.5 Flash model has been quite a success. However, the company paired it with a quietly nerfed AI Pro plan, and the tighter Gemini usage limits ended up frustrating users, especially those who used it for coding in Antigravity. Google reacted by increasing Antigravity’s limits by 9x (across two increases), but that still doesn’t seem enough. Now, Google has introduced a new Gemini model that uses even fewer tokens than Gemini 3.5 Flash.

Varun Mohan, Director at Google DeepMind, working on Antigravity, noted user concerns that Antigravity was using too many tokens for simple tasks. Consequently, Google has now introduced Gemini 3.5 Flash (Low) as a way to optimize token usage for these simple tasks.

The original Gemini 3.5 Flash now seems to have been renamed to Gemini 3.5 Flash (Medium). Google says the new Low variant generates around 45% fewer tokens than the Medium variant. Further, it also “generally” outperforms the older Gemini 3 Flash (Google is calling it “High” variant) on SWE tasks.

Alongside, Google has reset the Gemini quota across all Gemini plans, paid or free. This ensures users have all the tokens they need for their SWE work this week.

A user questioned whether Google did any “testing” before the launch, as it seems the company is testing on its users. To this, Mr. Mohan acknowledged the feeling but noted that the Google team has been using the model internally for a while. Google made Gemini 3.5 Flash good and fast at solving complex tasks, but ended up with a blind spot in measuring token usage for simpler tasks.

Mr. Mohan promises improvement in the future.

Expanding on the changes to Gemini 3.5 Flash (Low), Mr. Mohan notes that 3.5 Flash (Low) adjusts the model’s effort level but does not cut corners on the system prompt or context compaction.

The newer model is intended for simpler tasks. Users should use the higher-effort level for complex tasks.

It’s important to note that the answers were given in the context of Antigravity, Google’s AI-centric software development platform. I don’t see Gemini 3.5 Flash (Low) listed separately as a model within the Gemini app, and it’s unlikely to make its way here either. That said, I haven’t been impacted by high usage on simple tasks, so the problem could also be unique to Antigravity and other development-focused tools.