Google quietly added a third effort tier to its newest coding model on May 25, 2026: Gemini 3.5 Flash (Low). The variant generates roughly 45 percent fewer tokens than Flash (Medium) on the same prompts, and Varun Mohan, the Director leading Google's Antigravity platform, says it generally outperforms the legacy Gemini 3 Flash (now retroactively called High) on SWE-bench-style coding tasks. The original Flash you have been using in Antigravity since Google I/O has been silently relabelled Medium.
How to switch to Flash (Low) in Antigravity
Open the model picker in any Antigravity workspace and select Gemini 3.5 Flash (Low) instead of the default. Keep Medium for any task that involves multi-step planning, large refactors, or unfamiliar codebases. Use Low for the things that were eating your monthly quota: file renames, boilerplate scaffolding, single-function bug fixes, doc string generation, simple test stubs, and one-shot lint fixes. The system prompt and context compaction stay identical between tiers, so the only thing changing is the model's internal effort budget.
Why it matters
Antigravity users have been complaining for two weeks that the agent burns tokens on tasks that should be cheap. Mohan acknowledged the feedback directly when announcing Low, and Google has reset the Gemini quota across all paid and free plans alongside the launch. If you hit your token cap last week and stopped using Antigravity, your quota is now fresh and a cheaper tier is available for the work you do most. For studios paying per-seat, Low effectively expands the same dollar budget by about 80 percent on routine tasks.
Key details
Flash (Low) ships only inside Antigravity for now; it does not appear on the public Gemini 3.5 Flash model card at DeepMind, where Flash is still listed as a single Preview model. The tier swap means the variant you used yesterday is now Medium by default. Google has not committed to a public API endpoint for Low, so external IDE wrappers like Cline and Continue still call the Medium tier. Pricing on Gemini 3.5 Flash through Vertex AI and AI Studio is unchanged: the discount is purely a token-volume reduction, not a per-token cut. The agent-spawning workflow Mohan demonstrated at I/O 2026 still uses Medium for orchestration and now spawns Low workers for subtasks.
What to do next
Open Antigravity, set Flash (Low) as your default for routine work, and reserve Medium for anything you would have asked a senior engineer to plan. Track your token use for a week and compare against the prior week. Then file feedback in the Antigravity Discord on tasks where Low failed but Medium succeeded, since Google's tuning loop on the new tier is still active.
