Skip to content

lvbench: 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 340 66.2% 61.0–71.0 2 $0.099 28,348 4.08 14.3 s 44.2 s 39% 74% (n=336)
agent 2026-09-17 (claude-sonnet-5) 340 67.6% 62.5–72.4 1 $0.104 30,479 4.03 15.0 s 39.9 s 49% 74% (n=331)
agent 2026-09-17 (gemini-3.8-flash) 340 77.1% 72.3–81.2 0 $0.026 28,854 4.45 15.6 s 33.5 s 31% 74% (n=319)
retrieval-only 340 50.3% 45.0–55.6 1 $0.025 6,672 1.01 8.5 s 15.2 s 100% 58% (n=339)
uniform-32 (claude-sonnet-5) 340 52.1% 46.8–57.3 14 $0.037 11,718 0.00 11.8 s 22.3 s 100%
gemini-3.8-flash agentic video 338 80.8% 76.2–84.6 14 $0.059 39,531 3.40 13.2 s 95.6 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
entity recognition 65% (n=139) 69% (n=139) 78% (n=139) 45% (n=139) 50% (n=139) 78% (n=138)
event understanding 64% (n=135) 69% (n=135) 76% (n=135) 50% (n=135) 57% (n=135) 81% (n=135)
key information retrieval 76% (n=74) 70% (n=74) 82% (n=74) 57% (n=74) 45% (n=74) 88% (n=73)
reasoning 61% (n=38) 53% (n=38) 61% (n=38) 45% (n=38) 55% (n=38) 67% (n=36)
summarization 36% (n=14) 50% (n=14) 93% (n=14) 50% (n=14) 43% (n=14) 100% (n=14)
temporal grounding 78% (n=36) 78% (n=36) 92% (n=36) 56% (n=36) 53% (n=36) 92% (n=36)

Tool-call profiles

agent: search > search > view ×44; search > search ×30; search ×21; view ×16; search > view ×10; search > search > get_transcript ×8; search > search > view > view > view > view ×6; search > search > search > search > search > search ×6

agent 2026-09-17 (claude-sonnet-5): search ×17; view ×16; search > find_mentions ×15; search > find_mentions > view ×15; search > search > view ×12; (none) ×8; search > search ×8; search > get_transcript ×8

agent 2026-09-17 (gemini-3.8-flash): view ×15; search > get_transcript ×10; search > get_transcript > view ×9; search > view ×9; search > get_transcript > get_ocr > view ×4; search > get_ocr > view ×4; get_transcript > view ×4; get_ocr > view ×4

retrieval-only: search ×340

uniform-32 (claude-sonnet-5): (none) ×340

gemini-3.8-flash agentic video: processing_call×2 ×109; processing_call×1 ×81; processing_call×3 ×50; processing_call×4 ×26; processing_call×5 ×20; processing_call×6 ×12; processing_call×7 ×10; processing_call×8 ×8

Runs

  • agent: /data/videoindex/eval/runs/lvbench-f0.25-s1-agent.json — {"benchmark": "lvbench", "policy": "agent", "budget_usd": 0.5, "budget_tokens": 120000, "max_tool_calls": 6, "sample": 387, "fraction": 0.25, "seed": 1, "skipped_not_indexed": 47, "vi_version": "vi 0.1.0", "started": "2026-09-14T00:47:44Z"}
  • agent 2026-09-17 (claude-sonnet-5): /data/videoindex/eval/runs/lvbench-f0.25-s1-agent-v2.json — {"benchmark": "lvbench", "policy": "agent", "model": null, "label": "agent 2026-09-17 (claude-sonnet-5)", "budget_usd": 0.5, "budget_tokens": 120000, "max_tool_calls": 6, "sample": 387, "fraction": 0.25, "seed": 1, "skipped_not_indexed": 47, "vi_version": "vi 0.1.0", "started": "2026-09-17T14:46:27Z"}
  • agent 2026-09-17 (gemini-3.8-flash): /data/videoindex/eval/runs/lvbench-f0.25-s1-agent-gemini.json — {"benchmark": "lvbench", "policy": "agent", "model": "gemini-3.8-flash", "budget_usd": 0.5, "budget_tokens": 120000, "max_tool_calls": 6, "sample": 387, "fraction": 0.25, "seed": 1, "skipped_not_indexed": 47, "vi_version": "vi 0.1.0", "started": "2026-09-17T13:51:41Z", "label": "agent 2026-09-17 (gemini-3.8-flash)"}
  • retrieval-only: /data/videoindex/eval/runs/lvbench-f0.25-s1-retrieval-only.json — {"benchmark": "lvbench", "policy": "retrieval-only", "budget_usd": 0.5, "budget_tokens": 120000, "max_tool_calls": 6, "sample": 387, "fraction": 0.25, "seed": 1, "skipped_not_indexed": 47, "vi_version": "vi 0.1.0", "started": "2026-09-14T00:50:02Z"}
  • uniform-32 (claude-sonnet-5): /data/videoindex/eval/runs/lvbench-f0.25-s1-uniform-32.json — {"benchmark": "lvbench", "policy": "uniform-32", "model": "claude-sonnet-5", "frames": 32, "transcript": true, "sample": 387, "fraction": 0.25, "seed": 1, "started": "2026-09-13T23:44:46Z"}
  • gemini-3.8-flash agentic video: /data/videoindex/eval/runs/lvbench-f0.25-s1-gemini-3.8-flash-agentic.json — {"benchmark": "lvbench", "policy": "gemini-agentic", "model": "gemini-3.8-flash", "mode": "agentic", "thinking": null, "sample": 387, "fraction": 0.25, "seed": 1, "started": "2026-09-14T03:56:39Z"}

Questions per run: up to 340. Benchmark videos are public YouTube content and may be in model training data; read per-configuration deltas rather than absolute scores.