Decoding 23,515 Antigravity SQLite Records: ~49.5M Gemini Tokens Billed With 0% KV Cache
fatrabbit29 · reddit · 2026-11-21
A Google One AI Ultra subscriber decoded 23,515 local Antigravity SQLite protobuf records (Nov 12–20) and found:
- 0% KV cache regression: 49.5M uncached Gemini 4 input tokens billed against paid quota on same-model <45s follow-ups; a Claude Opus/Sonnet control group in the same app had zero failures across 623 turns.
- Undisclosed backend swaps: UI shows 'Gemini 4 Hydrogen (High)' while protobuf backend IDs reveal silent '-tiered' and 'flash-worker' routing with 50–60% lower average thinking tokens.
- 64.9% zero-thinking spikes: on Nov 19, most calls returned thinkingtokens=0, median zero.
- Failed empty responses still consumed quota until an HTTP 429 RESOURCEEXHAUSTED lockout.
- Support initially offered a refund catch-22 that would cancel bundled 30TB storage; after raw Response IDs were submitted, Tier-2 offered a refund without cancelling the plan.
A zero-dependency read-only audit script lets others check their own local SQLite DBs in seconds.
More from Infra
- MLX over Vulkan hits 124 tok/s on RTX 5060 Ti, beating llama.cpp CUDA by 25% — SumitGup · 2026-11-22
- Intruder swapped Prime Intellect's hosted DeepSeek for an abliterated copy for 53 hours — mawcopolow · 2026-11-22
- Token economics model projects $1.14T global API spend by 2030, but only 1% of users pay beyond coding — AccBalanced · 2026-11-22
- Zero-degradation Strata tuning: fixing layer_split lifts decode ~10-15% and halves CPU expert misses — PathfinderTactician · 2026-11-22
- llama.cpp Vulkan SCALE fix PR #32735 delivers 2.3-3.2x speedup on Radeon R9700 — pmttyji · 2026-11-22
- $1400 decade-old V100 32GB runs Qwen 27B: Q4_K_M +48% gen speed, Q8_0 +16% prefill — PuzzleheadedArm3732 · 2026-11-22