Multimodal · Benchmark profile

LVBench

A long-video understanding benchmark for retrieving and reasoning over information distributed across extended video inputs.

ReferenceAugust 2026Active4 models

Data verified 13 Aug 2026 · Methodology 2.2.0

Benchmark score on LVBench

Leaderboard · 4 models

Gemini 3.7 Flash leads at 85.4%, followed by Gemini 3.6 Flash (84.2%) and GPT-5.6 Terra (78.9%).

Display onlyMultimodalAugust 2026

#1Closed

Gemini 3.7 Flash

Google

gemini-3.7-flash

85.4%

Overall #6 · Context 1.05M tokens

#2Closed

Gemini 3.6 Flash

Google

gemini-3.6-flash

84.2%

Overall #20 · Context 1.05M tokens

#3Closed

GPT-5.6 Terra

OpenAI

gpt-5.6-terra

78.9%

Overall #9 · Context 1.05M tokens

ViewModelsColours

GoogleOpenAIAnthropic

Lumina

LVBench leaderboard

Accessible data for LVBench leaderboard
Model	Rank	Model ID	Provider	Score	Unit
Gemini 3.7 Flash	#1	gemini-3-7-flash	Google	85.4	percent
Gemini 3.6 Flash	#2	gemini-3-6-flash	Google	84.2	percent
GPT-5.6 Terra	#3	gpt-5-6-terra	OpenAI	78.9	percent
Claude Sonnet 5	#4	claude-sonnet-5	Anthropic	68.5	percent

Benchmark score table (4 models)

One best compatible score per canonical product · higher is better

LVBench leaderboard ranked by score
Rank	Model	Provider	License	Evidence use	Score
#1	Gemini 3.7 Flash gemini-3.7-flash	Google	closed	Reference only	85.4%
#2	Gemini 3.6 Flash gemini-3.6-flash	Google	closed	Reference only	84.2%
#3	GPT-5.6 Terra gpt-5.6-terra	OpenAI	closed	Reference only	78.9%
#4	Claude Sonnet 5 claude-sonnet-5	Anthropic	closed	Reference only	68.5%

The top of this snapshot is led by Gemini 3.7 Flash at 85.4%; third place is 6.5 points behind. The top-4 spread is 16.9 points.

Configuration-preserving evidencePublished source rows (4)Exact source labels, configurations, versions and dated observations. Same-snapshot canonical/configuration projections are shown once.

Every distinct dated source-published row for LVBench, including exact model configuration labels
Exact published model / configuration	Canonical product	Score	Version / system	Evidence	Source / observed
Gemini 3.7 Flash gemini-3-7-flash-medium	Gemini 3.7 Flash	85.4%	August 2026LVBench self-computed; 1024 frames for Gemini/GPT and 300 for Sonnet	Reference only provider-reported	Google DeepMind permanent refresh source ↗Observed 2026-08-13 · checked 2026-08-13
Gemini 3.6 Flash gemini-3-6-flash-high	Gemini 3.6 Flash	84.2%	August 2026LVBench self-computed; 1024 frames for Gemini/GPT and 300 for Sonnet	Reference only provider-reported	Google DeepMind permanent refresh source ↗Observed 2026-08-13 · checked 2026-08-13
GPT-5.6 Terra gpt-5-6-terra-max	GPT-5.6 Terra	78.9%	August 2026LVBench self-computed; 1024 frames for Gemini/GPT and 300 for Sonnet	Reference only provider-reported	Google DeepMind permanent refresh source ↗Observed 2026-08-13 · checked 2026-08-13
Claude Sonnet 5 claude-sonnet-5-max	Claude Sonnet 5	68.5%	August 2026LVBench self-computed; 1024 frames for Gemini/GPT and 300 for Sonnet	Reference only provider-reported	Google DeepMind permanent refresh source ↗Observed 2026-08-13 · checked 2026-08-13

Compare top models

Gemini 3.7 Flash vs Gemini 3.6 Flash Gemini 3.6 Flash vs GPT-5.6 Terra GPT-5.6 Terra vs Claude Sonnet 5

About LVBench

Definition and scoring

Organisation: LVBench
Category: Multimodal
Version: August 2026
Direction: higher is better
Ranking use: Reference
Contamination risk: Unknown

A long-video understanding benchmark for retrieving and reasoning over information distributed across extended video inputs. Every genuine source row stays tied to its exact model label, configuration, benchmark version and evaluation system. The summary chart shows one best compatible score per canonical product; Score 2.0 admits only explicitly mapped, frozen protocols and keeps incompatible configurations separate.

Open benchmark source ↗

FAQ

Quick answers

What does LVBench measure?

A long-video understanding benchmark for retrieving and reasoning over information distributed across extended video inputs.

Which model scores highest on LVBench?

Gemini 3.7 Flash by Google currently leads with 85.4%.

How many models are evaluated on LVBench?

4 canonical model products have a qualifying score, backed by 4 distinct dated source rows.

Does this affect overall Lumina rank?

No. This benchmark is display-only and does not enter the overall Lumina composite.

Multimodal · Benchmark profile

LVBench

A long-video understanding benchmark for retrieving and reasoning over information distributed across extended video inputs.

ReferenceAugust 2026Active4 models

Data verified 13 Aug 2026 · Methodology 2.2.0

Benchmark score on LVBench