Hi everyone,
I am writing to report a highly inefficient behavior regarding the new Gemini 3.5 Flash (High) model when running within the Antigravity IDE. I am currently on the Google AI Pro ($20/month) plan, and this issue is making agentic workflows financially and operationally unviable.
1. The Bug: Infinite Redundant Context Ingestion Loops
When executing agentic tasks that require background knowledge or raw codebase context, Gemini 3.5 Flash (High) falls into a severe “reading loop.” It repeatedly re-ingests the exact same files or context blocks up to 3 or 4 times within a single execution cycle before attempting any actual output action.
This aligns heavily with what other users have described as “Context Window Compaction Amnesia” in this forum (see: Antigravity 2.0 lacklustre? Gemini 3.5 Flash seems). The platform triggers an internal compaction, the model loses the precise file tracking, and its immediate next tool-call is to re-read everything from scratch, exponentially multiplying the input costs.
2. Cross-Model Validation (Control Group)
To confirm this isn’t a prompt engineering issue on my end, I deployed the exact same agentic loop architecture, XML-delimited background files, and system instructions across competing models:
- Flawless execution (Linear token footprint): Gemini 3.1 Pro, Anthropic Claude 4.6 Sonnet, and Claude 4.6 Opus.
- Failed execution (Massive quota drain / looping): Gemini 3.5 Flash (High).
3. Total Metric Opacity in the IDE
To make matters worse, the Antigravity User Settings (“Models” tab) completely obscures granular token counts. There is no way to audit the exact promptTokenCount or the hidden thinking tokens being billed to our accounts.
The only indicator we get is watching our daily/weekly quota percentages vanish into thin air (e.g., losing a massive chunk of our tier allocation from just a single loop iteration) without a graceful warning threshold or checkpoint.
Environment Details:
- IDE: Antigravity IDE (VSCode style desktop version)
- Subscription Plan: Google AI Pro ($20)
- Faulty Model: Gemini 3.5 Flash (High)
Looking for answers:
Is Google currently tracking this interaction between Gemini 3.5 Flash’s internal agent reasoning and Antigravity’s workspace indexing? Also, can we please get transparent, granular token logging (usageMetadata) natively inside the IDE so we can properly debug our agent costs?
Thanks!