TwelveLabs
Listed at https://twelvelabs.io
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | Free plan: 600 minutes video indexing + API usage (Pegasus 1.2 $0.042/min indexing one-time, $0.021/min input video, $0.0075/1k tokens output text, $0.0015/min monthly embedding infrastructure). Developer plan: paid overage, 90-day index retention lifted to unlimited. | per million tokens |
| Output | Same per-minute and per-1k-tokens billing. Marengo embeddings billed separately by indexed minute. | per million tokens |
🤖 Supported Models (3)
✨ Pros
- ✓Leading dedicated video-understanding API: Pegasus analyzes visual frames + temporal sequence + speech + on-screen text together and returns natural-language descriptions
- ✓Marengo 3.0 is a 27B-parameter video foundation model producing multimodal embeddings (frame / audio / speech / text) that power any-to-any video retrieval
- ✓RESTful API + official Python / Node / Go / Java SDKs; OpenAPI 3 spec; asynchronous indexing task model (POST to create task → poll for status → fetch result)
- ✓Developer-friendly Playground and MCP server integration so AI clients like Claude Code and Cursor can call TwelveLabs directly
- ✓Knowledge Stores + Agents (Research Preview): upload videos → automatic chunking → cross-video Q&A and retrieval
- ✓$100M Series B in 2026, valued as a unicorn — one of the top-funded pure-play video-AI startups
- ✓Free plan grants 600 cumulative minutes of indexing — enough to fully PoC a small or mid-size video understanding project without a credit card
⚠️ Cons
- ×Per-minute video billing (not per-token); long-form content (surveillance / sports / full movies) needs a different cost model than text LLM APIs
- ×Indexing tasks are async polling — typical 30 seconds to several minutes for results, so not suited for real-time generation use cases
- ×No mainland China direct endpoint; production deployment from China requires a proxy or overseas nodes
- ×Pegasus 1.5 vs 1.2 mainly improves multimodal fusion quality; on pure-text or pure-frame tasks the marginal gain is small
- ×Knowledge Store and Agents are still in Research Preview; API surface may change
- ×No function calling / tool use — pure video-understanding API; downstream reasoning requires an LLM (OpenAI / Claude / Kimi) in the orchestration loop
- ×No affiliate program
🎯 Best For
Teams that need semantic understanding / search / structuring of video archives (media archives, surveillance, sports footage, online courses, corporate training video); RAG systems that need long-video context for an LLM (use Marengo to retrieve relevant clips, then feed them to Claude / GPT); ad and e-commerce video content moderation (joint visual + speech judgment); complements Kling / Luma / Runway generation APIs — those generate, TwelveLabs understands
💰 Pricing & Plans
| Component | Free Plan | Developer Plan | Notes |
|---|---|---|---|
| Video indexing (one-time, per minute) | $0.042/min | $0.042/min | Pegasus 1.2 / 1.5 — charged once when the video is indexed |
| Input video (per minute, per call) | $0.021/min | $0.021/min | Pegasus 1.2 — Analyze & Segment, Search, RAG calls |
| Output text (per 1k tokens) | $0.0075/1k | $0.0075/1k | Pegasus 1.2 — natural-language generation |
| Embedding infrastructure (monthly) | $0.0015/min | $0.0015/min | Billed monthly by total indexed minutes stored |
| Marengo 3.0 embeddings | Per-minute | Per-minute | Same per-indexed-minute billing; used by Search / RAG |
| Index retention | 90 days | Unlimited | Free indexes expire after 90 days; Developer keeps them |
| Monthly subscription fee | $0 | $0 | Developer plan is pay-as-you-go only; no monthly minimum |
| Free signup credit | 600 minutes (cumulative) | — | Enough to PoC a mid-size video understanding project |
| Enterprise contracts | — | Available on request | Volume discounts, custom rate limits, private deployment |
🔧 API & Developer Experience
- •API Surface: RESTful JSON API at https://api.twelvelabs.io/v1.3. Standard verbs (GET / POST / PUT / DELETE). Auth via x-api-key header. Returns X-Api-Version on every response. Compatible with any language that speaks HTTP; official SDKs remove boilerplate.
- •Official SDKs: Python (twelvelabs-python on PyPI), Node.js, Go, Java, plus a community Ruby SDK. All four first-party SDKs are thin wrappers around the same REST surface and include example notebooks for indexing, search, and embedding tasks.
- •Indexing Workflow: Asynchronous task model: POST /v1.3/tasks to upload a video (or reference an existing asset) → poll GET /v1.3/tasks/{task_id} until status is 'ready' → call Search / Embed / Analyze on the indexed video. Typical indexing latency: 30s for a 5-minute clip; several minutes for hour-long content.
- •Endpoints of Note: POST /tasks (create indexing task), POST /search (any-to-any search across indexes), POST /embed (generate Marengo embeddings), POST /analyze (run Pegasus on a video), POST /gist (Pegasus-generated natural-language summary), POST /knowledge-stores (cross-video Q&A in research preview).
- •MCP Server: TwelveLabs exposes an MCP server at https://docs.twelvelabs.io/_mcp/server so AI agents like Claude Code and Cursor can call Search / Analyze / Embed directly from a chat session — useful for prototyping RAG pipelines before writing integration code.
- •Rate Limits: Rate limits scale with monthly spend (documented in API docs as 'Rate limits apply and scale with monthly spend'). Free plan: low concurrency cap. Developer plan: higher concurrency. Enterprise: negotiated. 429 returned with Retry-After header when exceeded.
- •Error Handling: Standard HTTP codes (200 / 201 / 400 / 401 / 404 / 429). Errors include machine-readable 'code' field and human-readable 'message'. SDKs raise typed exceptions with these fields so callers can branch on error class without parsing strings.
🎥 Video Understanding Capabilities
TwelveLabs is the leading dedicated video-understanding API in 2026, and its flagship Pegasus model is the only production model that fuses visual frames, temporal sequence, speech audio, and on-screen text into a single natural-language description. A Pegasus call returns a coherent gist: setting, speakers, actions, on-screen text — tied together with causal language. Pegasus 1.5 improves multimodal fusion over 1.2 by jointly encoding all four streams in one transformer pass rather than stitching them post-hoc, which is the difference between 'describe the scene' and 'describe what is happening and why.' For retrieval, Marengo 3.0 — a 27B-parameter video foundation model — produces multimodal embeddings that capture visual appearance and acoustic context together, so text queries like 'the moment the speaker mentions pricing' return the right clip.
🌐 Regional Availability & Latency
TwelveLabs is a US-based company (San Francisco) hosting the public API on AWS us-west-2 and us-east-1. There is no mainland China regional endpoint or official China access program, so production traffic from China must either proxy through overseas nodes or run TwelveLabs on a private deployment in an AWS region of the customer's choice (Enterprise contracts only). Latency for indexing a 5-minute clip is typically 30 seconds; for an hour-long video, several minutes. Search and Analyze calls are synchronous over HTTP and add 1-3 seconds of network round-trip to model inference time. The async indexing model means a 'submit and poll' workflow rather than request-response, which suits batch pipelines but not real-time video generation (use a generation API like Kling or Luma for that).