Runway

💰 Pricing & Plans

🔧 API & Developer Experience

• •Asynchronous task lifecycle on every generation: A generation POST returns a task id, not a file; poll GET /v1/tasks/{id} until status is SUCCEEDED or FAILED. THROTTLED is a queue state rather than an error, DELETE /v1/tasks/{id} cancels, and the SDK waitForTaskOutput() handles the backoff and raises a distinct TaskFailedError
• •One required version header: Every request must carry X-Runway-Version: 2024-11-06 next to the Bearer key. The header pins ratio and duration semantics, so swapping models without changing the version string stays safe
• •Official Node and Python SDKs: @runwayml/sdk and the runwayml PyPI package read RUNWAYML_API_SECRET, inject the version header and expose typed methods per endpoint family; base URL is https://api.dev.runwayml.com
• •MCP for agent-driven media: dev.runwayml.com/mcp exposes task lookup, model router management and docs to Codex, Cursor, Claude Code and Grok Bot without pasting an API key into mcp.json; mcp.runwayml.com and the GitHub server cover in-chat generation
• •Model Router: Save a routing configuration once and call it by configId instead of hardcoding a model; the router picks by cost, latency or quality, honours allow and deny lists plus per-modality credit ceilings, and dryRun previews the decision
• •Recipes and published Workflows: Seven recipes (product_ad, product_swap, product_ugc, multi_shot_video, ad_localization, marketing_stock_image, product_campaign_image) wrap multi-step generation behind one call; Workflows are node-graph pipelines built in the app and published as versioned endpoints polled at /v1/workflow_invocations/{id}
• •Tier-based concurrency and no RPM cap: Concurrency scales from 1 to 20 tasks by tier with automatic upgrades after $50, $100, $1,000 and $5,000 of purchases, and there is no requests-per-minute limit: excess requests queue server-side, so no rate limiter is needed in your integration

🎬 One Ledger, Sixty Models, Ten Labs

Runway Dev is no longer just Runway's own video models. A single credit balance and a single API now front sixty media models spanning Runway, Kuaishou, ByteDance, Google, xAI, MiniMax, Alibaba, OpenAI, Ideogram and Bria: twenty-seven video, twenty image, eight audio, four upscale or HDR, one real-time. That breadth changes the integration problem. The hard part is no longer model access but per-model parameter drift: Gen-4.5 accepts two ratios and a two-to-ten second duration, Seedance 2.5 accepts twenty-four ratios up to 4K with a 3,500-character prompt, and each carries its own audio, reference and draft-mode rules, because every generation request body is a discriminated union keyed on model. Runway's answer is the Model Router, which swaps a hardcoded identifier for a saved policy over cost, latency or quality. The contract still moves per model, so read the spec before shipping. Recipes and Workflows then let a multi-shot ad pipeline look like one endpoint.

🌐 China Access & Latency

Runway Dev has no mainland China region and no Chinese data residency, so mainland access needs a proxy; the endpoint is not blocked outright, but sustained throughput is unreliable without one. From an Alibaba Cloud Malaysia host the edge is quick: api.dev.runwayml.com sits behind CloudFront and was served from the KUL51 Kuala Lumpur POP, with a median time to first byte of 489 ms across eight samples and a 253 to 750 ms range. A mainland client tunnelled through Hong Kong or Singapore lands in the same APAC edge pool and usually sees comparable numbers. The saving grace is that generation is asynchronous: you post a job, poll for the result, and a mainland user was waiting on a task that takes seconds to minutes anyway. Latency therefore only bites when uploading input media, where large source clips dominate the round trip rather than the API handshake. Teams with data-residency requirements should keep assets on their own side, generate on Runway, and pull the temporary output URLs, which expire within 24 to 48 hours, back into internal storage.