pixelml-av + Grok + Jev: Agentic Video Understanding 100x Cheaper and 10x Faster Than Gemini 3.8 Flash
Sean Phan posted a 65-second slide walkthrough: on a 75-minute video with 8 questions, the open-source pixelml-av library (av) paired with Grok and Jev costs $0.0022 per query at 4.0 s, versus $0.2253 per query at 39.9 s for native Gemini 3.8 Flash over the full video, a claimed ~100x cheaper and ~10x faster. The deck itemizes both sides: the native path spent 2,046,870 tokens and $1.1242 of query-model cost across 8 saved attempts, while the AV path spent 25,964 tokens and $0.0111 query-model cost per query plus a one-time $0.57 prep (Grok captioning sampled frames, Flash Lite transcribing audio once, index reusable). The pipeline runs in six stages: question, retrieve candidate scenes from the local index, Jev relevance filtering, local bounded context (temporal grouping assembles a small evidence window with no extra LLM summary call), Grok answering with citations, and a Jev support check; weak or unknown support can trigger a bounded sampled-still inspection of selected frames (not used in this run). av is the open-source video-indexing library behind PixelML Composer/Sentinel agents, and the linked cookbook recipe jev-refined-ask documents the av ingest / av ask CLI, the –no-refine and AV_REFINE_ENABLED=false opt-outs, an external transcript import seam and a budget-ledgered Gemini transcription helper. The author frames this as a new opportunity for enterprises with large video corpora and many queries over the same corpus.
