Skip to content

minerva: 25% stratified sample, seed 1

Generated 2026-09-17 by eval/report.py from 6 run file(s). Accuracy is exact match on the option letter; intervals are 95% Wilson. Costs are provider list prices per question; indexing cost is not included for the index-based configurations.

Configuration n accuracy 95% CI unparsed cost / q tokens / q tool calls / q latency p50 latency p95 no-decode citation in range
agent 310 51.9% 46.4–57.4 22 $0.150 40,830 4.84 24.2 s 76.2 s 14%
agent 2026-09-17 (claude-sonnet-5) 310 55.8% 50.2–61.2 6 $0.151 41,419 4.56 23.0 s 92.5 s 21%
agent 2026-09-17 (gemini-3.8-flash) 310 66.8% 61.4–71.8 8 $0.035 33,579 4.69 22.1 s 47.0 s 8%
retrieval-only 310 37.1% 31.9–42.6 11 $0.034 8,624 1.10 11.0 s 33.0 s 100%
uniform-32 (claude-sonnet-5) 310 41.3% 35.9–46.8 35 $0.021 5,802 0.00 13.7 s 22.7 s 100%
gemini-3.8-flash agentic video 310 76.5% 71.4–80.8 6 $0.051 45,673 4.08 16.2 s 134.9 s 100%

Where this stands against Gemini agentic video

The gemini-… row is Google's Gemini 3.8 Flash with processing: "agentic" asked the same questions over the same video files through the Interactions API (run once and cached; it is not re-run with every matrix). Google's own announcement ("Introducing agentic video in Gemini", 2026-09-01) reports agentic mode against static whole-video processing on LongVideoBench as up to 88% fewer tokens, up to 66% lower cost and up to 7% higher accuracy, without absolute scores; the numbers here are absolute, on LVBench, and directly comparable across rows because every row answered the same questions under the same scoring. VideoIndex's per-question cost excludes indexing (done once per video); Gemini's per-question cost is the whole cost. Both systems are scored with the same letter parser.

Accuracy by task type

Task type agent agent 2026-09-17 (claude-sonnet-5) agent 2026-09-17 (gemini-3.8-flash) retrieval-only uniform-32 (claude-sonnet-5) gemini-3.8-flash agentic video
Cause and Effect 55% (n=11) 36% (n=11) 82% (n=11) 64% (n=11) 45% (n=11) 91% (n=11)
Counterfactual 12% (n=8) 38% (n=8) 75% (n=8) 0% (n=8) 25% (n=8) 75% (n=8)
Counting 38% (n=63) 38% (n=63) 49% (n=63) 21% (n=63) 30% (n=63) 62% (n=63)
Event Occurence 49% (n=45) 56% (n=45) 60% (n=45) 40% (n=45) 38% (n=45) 78% (n=45)
Goal Reasoning 67% (n=3) 67% (n=3) 100% (n=3) 67% (n=3) 33% (n=3) 67% (n=3)
Listening 66% (n=32) 75% (n=32) 78% (n=32) 56% (n=32) 62% (n=32) 88% (n=32)
Numerical Reasoning 58% (n=12) 75% (n=12) 75% (n=12) 33% (n=12) 33% (n=12) 92% (n=12)
Object Recognition 64% (n=56) 61% (n=56) 75% (n=56) 45% (n=56) 54% (n=56) 79% (n=56)
Reading 72% (n=29) 72% (n=29) 79% (n=29) 38% (n=29) 41% (n=29) 83% (n=29)
Situational Awareness 57% (n=7) 86% (n=7) 71% (n=7) 57% (n=7) 29% (n=7) 86% (n=7)
Spatial Perception 36% (n=11) 45% (n=11) 55% (n=11) 36% (n=11) 27% (n=11) 73% (n=11)
State Changes 67% (n=3) 33% (n=3) 100% (n=3) 0% (n=3) 67% (n=3) 67% (n=3)
Temporal Reasoning 37% (n=30) 50% (n=30) 60% (n=30) 30% (n=30) 37% (n=30) 73% (n=30)

Tool-call profiles

agent: search > search > view ×35; view ×22; search > view ×10; search > search ×10; search > search > view > view > view > view ×10; view > view ×7; search > search > view > view ×7; search > search > search > search > search > view ×4

agent 2026-09-17 (claude-sonnet-5): view ×27; search > view ×16; search > search > view ×8; view > view ×7; search > search > view > view > view > view ×6; search > get_transcript > view ×5; search > find_mentions > view ×5; search > get_ocr ×5

agent 2026-09-17 (gemini-3.8-flash): view ×19; search > get_transcript > view > view > view > view ×7; search > view > view > view > view > view ×7; get_ocr > view ×6; search > get_ocr > view ×6; view > view ×5; view > view > view ×4; get_transcript > view ×4

retrieval-only: search ×310

uniform-32 (claude-sonnet-5): (none) ×310

gemini-3.8-flash agentic video: processing_call×2 ×78; processing_call×1 ×55; processing_call×3 ×46; processing_call×4 ×40; processing_call×6 ×26; processing_call×5 ×18; processing_call×9 ×10; processing_call×7 ×9

Runs

  • agent: /data/videoindex/eval/runs/minerva-f0.25-s1-agent.json — {"benchmark": "minerva", "policy": "agent", "budget_usd": 0.5, "budget_tokens": 120000, "max_tool_calls": 6, "sample": 323, "fraction": 0.25, "seed": 1, "skipped_not_indexed": 13, "vi_version": "vi 0.1.0", "started": "2026-09-14T00:31:11Z"}
  • agent 2026-09-17 (claude-sonnet-5): /data/videoindex/eval/runs/minerva-f0.25-s1-agent-v2.json — {"benchmark": "minerva", "policy": "agent", "model": null, "label": "agent 2026-09-17 (claude-sonnet-5)", "budget_usd": 0.5, "budget_tokens": 120000, "max_tool_calls": 6, "sample": 323, "fraction": 0.25, "seed": 1, "skipped_not_indexed": 13, "vi_version": "vi 0.1.0", "started": "2026-09-17T15:11:52Z"}
  • agent 2026-09-17 (gemini-3.8-flash): /data/videoindex/eval/runs/minerva-f0.25-s1-agent-gemini.json — {"benchmark": "minerva", "policy": "agent", "model": "gemini-3.8-flash", "label": "agent 2026-09-17 (gemini-3.8-flash)", "budget_usd": 0.5, "budget_tokens": 120000, "max_tool_calls": 6, "sample": 323, "fraction": 0.25, "seed": 1, "skipped_not_indexed": 13, "vi_version": "vi 0.1.0", "started": "2026-09-17T14:15:52Z"}
  • retrieval-only: /data/videoindex/eval/runs/minerva-f0.25-s1-retrieval-only.json — {"benchmark": "minerva", "policy": "retrieval-only", "budget_usd": 0.5, "budget_tokens": 120000, "max_tool_calls": 6, "sample": 323, "fraction": 0.25, "seed": 1, "skipped_not_indexed": 13, "vi_version": "vi 0.1.0", "started": "2026-09-14T00:44:31Z"}
  • uniform-32 (claude-sonnet-5): /data/videoindex/eval/runs/minerva-f0.25-s1-uniform-32.json — {"benchmark": "minerva", "policy": "uniform-32", "model": "claude-sonnet-5", "frames": 32, "transcript": true, "sample": 323, "fraction": 0.25, "seed": 1, "started": "2026-09-13T21:06:31Z"}
  • gemini-3.8-flash agentic video: /data/videoindex/eval/runs/minerva-f0.25-s1-gemini-3.8-flash-agentic.json — {"benchmark": "minerva", "policy": "gemini-agentic", "model": "gemini-3.8-flash", "mode": "agentic", "thinking": null, "sample": 323, "fraction": 0.25, "seed": 1, "started": "2026-09-13T23:29:17Z"}

Questions per run: up to 310. Benchmark videos are public YouTube content and may be in model training data; read per-configuration deltas rather than absolute scores.