A video-native AI platform that provides APIs for indexing, searching, and analyzing video at scale. Unlike frame-by-frame or transcription-only tools, it understands visual content, audio, speech, and temporal relationships together. Two core models power it: Marengo (encoder) generates multimodal embeddings for any-to-any retrieval across video, audio, images, and text; Pegasus (reasoning model) turns video into summaries, structured metadata, and descriptions with timestamps. Key capabilities include natural-language semantic search across entire libraries, automatic segmentation, compliance/brand-safety review with explainable output, and highlight assembly. It processes roughly 60x real-time (an hour of video indexed in about a minute) and ships as REST API with Python/Node/Go/Java SDKs, MCP integrations, and deployment on AWS Bedrock. SOC 2 Type II certified. Pricing is per-minute: Free tier includes 600 minutes of indexing (90-day index retention, no credit card); Developer pay-as-you-go starts at $0.042/minute for indexing and $4 per 1,000 search queries. The company raised a $100M Series B in 2026 and is valued as a unicorn.
**Pros & Cons**:
Pros:
Purpose-built video understanding API, not a general multimodal model adapted for video—handles visual frames, temporal sequence, speech, and on-screen text together; Marengo produces shared embeddings across modalities, enabling cross-modal search (text query → find video moment, audio clip, or image); Pegasus generates timestamped descriptions and structured metadata, which downstream LLMs can consume; ~60x real-time indexing handles large archives (10,000+ hours per day on Developer tier); MCP server and SDKs make it callable from Claude, Cursor, and custom pipelines; SOC 2 Type II and air-gapped deployment options suit regulated industries; Free tier offers 600 minutes of indexing without a credit card, enough for a real proof of concept.
Cons: **API-only, no end-user UI**—there is no timeline to scrub or clip shelf to click; your team must write code or use an integration partner; **per-minute billing can scale quickly** for long-form archives (surveillance, sports, full movies), and storage plus indexing plus Analyze input add up; indexing is asynchronous (polling model), so it is not suited for real-time generation use cases; **model lineup moves fast** (Marengo, Pegasus, Jockey rev on their own schedule), so evaluations should be pinned to a dated test set; some user reports describe it as \"doesn\'t work\" and note a lack of independent third-party demos ; it is enterprise-first and not suitable for casual users or simple filename searches.
**Who is it for?**:Media and entertainment companies sitting on large video archives that need semantic search and scene-level retrieval; adtech and ecommerce teams building contextual targeting or video content moderation (joint visual + speech judgment); sports rights holders and news organizations pulling specific moments from footage; developers building RAG pipelines that need long-video context for an LLM (use Marengo to retrieve clips, then feed to Claude/GPT); government or security teams needing air-gapped video analysis.