After noticing rapid allowance drain and erratic reasoning starting Nov 15 on a $199.99/mo Google One AI Ultra (20x) plan, I decoded the raw gen_metadata protobuf billing records across all 373 local SQLite conversation DBs (~/.gemini/antigravity/conversations/*.db, 23,515 model calls).
Three verifiable findings stood out in Google’s own protobuf usage records:
Server-side 0% KV Cache bug (~49.5M uncached tokens overbilled): On consecutive same-model turns <25s apart with zero compaction, Gemini 4 repeatedly dropped cached_input_tokens to 0 (e.g., 945k uncached tokens billed across two turns 25s apart on Carbon High; 3.26M uncached tokens across 150s on Orchestrator). Meanwhile, hosted Claude Opus/Sonnet 5.5 inside the exact same Antigravity client had 0 (0.0%) zero-cache failures across 623 strict follow-ups.
Undisclosed “-tiered” & worker backend model IDs: Comparing field_21 (UI display label) vs field_19 (backend model string) shows Gemini 4 Hydrogen (High) silently switched on Nov 17 to “gemini-4-hydrogen-tiered-high” (with >60% lower mean thinking tokens), while 6,699 “Gemini 4 Hydrogen (Orchestrator)” calls ran on “gemini-flash-worker-agy”. Argon (Medium) calls returning thinking_tokens = 0 also jumped from 0.3% (Nov 12-13) to 64.9% (Nov 19).
PSA on Google One’s Bundled 30 TB Billing Catch-22: Because Google bundles developer AI Ultra compute inside Google One storage plans, standard Tier-1 courtesy refunds immediately cancel the entire membership and revoke your 30 TB Drive/Gmail storage (to the Tier-1 rep’s credit, they warned me before clicking it—otherwise my 3.5 TB Drive & Gmail would have been frozen immediately). Only after declining cancellation and uploading raw SQLite protobuf traces + server ResponseIDs did Tier-2 acknowledge an option to refund the last plan charge without cancelling storage (still pending manual review with no SLA).
Full write-up, redacted support transcript, and zero-dependency Python self-audit script so any Pro/Ultra subscriber can check their own local SQLite DBs in 5 seconds:
Sharing an updated v1.1 audit (audit_antigravity_sqlite.py now enforces strict ResponseID deduplication across forked/Swarm DBs, requires real per-step transcript timestamps on both turns, and verifies usage.3 == usage.9 + usage.10 with 24,319/24,319 exact matches):
Deduplicated Nov 15–20 KV Cache Summary:50,650,659 overbilled uncached tokens across 397 full 0%-cache wipes within <=300s (28,028,445 within <=45s), vs 0 / 648 (0.00%) for hosted Claude in the same client. Clean <30s single-model traces with zero intervening errors or timeouts:
Post-Hotfix Reproducible Leaks Still Active on Nov 19–20 (925 turns):
Client Harness Bug: Async SYSTEM_MESSAGE wakeups truncate gemini-flash-worker-agy cached prefix back to ~25K–27K within 5s–84s (51.1% in 486-turn session 80ccfd72; 100%3/3 in 611de086).
Orchestrator (Neon) Cache Thrashing:71 planner/worker switches incurred 15.55M uncached input (90.3% planner miss rate after >300s worker execution), causing Orchestrator to consume 2.58x more uncached input than straight Carbon Medium.
180s Server Eviction Cliff + Sub-30s Wipes: Continuous Carbon Medium stayed <3% miss under 180s, jumped to 46.2% at 180–240s and 92.9% past 300s, plus sudden 27s resets (611de086NH8Aa8ixHM-Z****: 323,074 cached → 325,737 uncached / 0 cached).
Latency Impact: For >=100K contexts, 0%-cache wipes increased median TTFT (d11) 2.57x (5.75s → 14.76s).