Runway API Guide: Runway Dev Setup, Pricing and Limits

The Runway API is sold as Runway Dev, a pay-as-you-go platform kept apart from Runway's app subscriptions. One credit costs $0.01, there is no monthly fee, and you prepay at least $10 before the first call ^[1]. Runway's flagship Gen-4.5 bills $0.12 a second of video, so a five-second clip is 60 credits, or $0.60 by our arithmetic. The same key reaches third-party models such as Veo 3.1, Kling 4.0 and Seedance 2.5 at their own credit rates. Our verdict is simple. Runway Dev suits developers who want several video labs behind one prepaid bill and one task format. It is the wrong purchase for clips made by hand, because on self-serve accounts app plans and API credits never mix.

From sign-up to a working key

Runway Dev lives at dev.runway.com, with documentation at docs.dev.runwayml.com and requests sent to api.dev.runwayml.com; the old RunwayML name survives in those hostnames. Signing up is free; generating is not. There is no free tier, and each project needs a prepaid balance before its first request. Credits are bought per project in the developer portal, and a web-app subscription is neither required nor helpful, since its credits do not carry over. Each project holds its own keys, credits, usage tier, members and billing.

Roles matter on day one. An Admin can do everything, a Billing admin handles credits and autobilling but cannot create keys, and a Developer manages keys only. Invitations default to Admin, so change the role when you invite a contractor. Keys begin with key_ and are shown exactly once. They belong to the project rather than the person, so removing someone does not revoke a key they created. Disable keys separately whenever a team member leaves.

The official SDKs are @runwayml/sdk for Node (Node 18 or newer) and runwayml for Python (3.8 or newer). Both read the key from the RUNWAYML_API_SECRET environment variable and retry 429 and 5xx errors on their own. Plain cURL works too: send the key as a Bearer token and add the X-Runway-Version header, currently 2024-11-06, on every request. Runway keeps an old API version working for four months after a new one appears, so pin the header and read the changelog before each upgrade.

The per-second price list

Every video model on Runway Dev bills by the second of output, at a rate set by the model you name in the request, and several carry a minimum charge per generation. The table lists Runway's own models at list price, read from the Runway Dev pricing page in November 2026; third-party models follow in the next paragraphs.

Professional formats cost extra. ProRes and PNG sequences add 5 credits a second on Gen-4.5 and Aleph 2.0, and 10-bit or HDR profiles add 20 a second, or 40 above about 4 megapixels. All charges round up to the next credit, and a finished task reports its final cost, the figure worth logging per job. In November 2026 Gen-4.5 outputs 720p clips of 2 to 10 seconds, Aleph 2.0 edits footage of 2 to 30 seconds at up to 1080p, and Act-Two drives one character for up to 30 seconds. The gen3a_turbo and gen4_aleph identifiers were removed in mid-2026.

Third-party models bill from the same balance at their own rates. Veo 3.1 costs 40 credits a second with audio and 20 without. Kling 4.0 lists at 15.7 credits a second at 720p and 23.5 at 1080p, with 20% off through November 30, 2026. Seedance 2.5 runs 20, 30 or 68 credits a second at 480p, 720p or 1080p, plus a charge per second of input video and an 80-credit minimum. These are Runway Dev credits, a different currency from the credits those labs sell on their own sites.

Budget for retakes. A clip that fails your eye still costs its credits, so the real cost per usable clip sits above the list rate. By our arithmetic, an API credit at 1.0¢ beats a Standard plan credit at 2.4¢ billed monthly and a Pro credit at 1.56¢, and matches Max at 1.0¢ billed monthly. That only matters if you would script the work anyway, because plan credits cannot pay for API calls; the Runway pricing guide covers the app plans.

Sending a job and getting the file back

Generation is asynchronous. A POST to /v1/text_to_video or /v1/image_to_video returns a task id, and you poll that task until it reads SUCCEEDED, FAILED or CANCELLED. Image to video requires promptImage, a URL or base64 data URI that can carry a first and a last frame. Gen-4.5 takes prompts of up to 1,000 characters, and its text-to-video output comes only as 1280:720 or 720:1280. Poll every five seconds or slower, with jitter and backoff. The SDK helper waitForTaskOutput gives up after ten minutes, and a timeout does not cancel the task, which keeps running and keeps its cost.

Inputs carry their own rules. URLs must be HTTPS on a domain name rather than an IP address, and redirects are not followed. Images go up to 16MB by URL in JPEG, PNG or WebP (GIF is refused), video and audio up to 32MB, and ephemeral uploads up to 200MB for 24 hours. The output side holds the real trap. Result links expire within 24 to 48 hours ^[2], so copy every file into your own storage the moment a task succeeds, never just the URL.

Tiers, throttling and failed tasks

Limits are set per model and per project by usage tier. There is no requests-per-minute cap. Instead, each tier allows a maximum number of generations in a rolling 24-hour window, returning 429 Too Many Requests beyond it, and caps credit spending over 30 days. Tasks beyond your concurrency limit come back THROTTLED and wait in rough submission order, so treat them as pending rather than building your own queue ^[3]. Higher tiers arrive automatically as spending grows; guaranteed limits need an exception request.

Failures split two ways. When an error on Runway's side kills a generation, the credits come back automatically. Moderation is different: it scans inputs and outputs, a blocked generation appears as a FAILED task rather than an HTTP error, and it costs the same as a successful one. SAFETY.INPUT failures are not refunded ^[4] and should not be retried. Repeated moderated requests can suspend the account, and a sudden 401 may mean exactly that. A THIRD_PARTY.UNAVAILABLE code means another lab's model is down, so wait before trying again.

Router, Recipes, Workflows and Characters

The Model Router, launched in July 2026, picks a model per request by cost, quality or latency, inside allow and deny lists and a credit ceiling. A dry run returns the chosen model and an estimated cost without billing anything, and routed requests pay the chosen model's standard rate with no router fee. One catch: router endpoints take aspect ratios such as 16:9, while direct model endpoints want pixel ratios such as 1280:720. Recipes bundle model choice, prompting and post-processing into one call for seven jobs, from product ads and product swaps to ad localization. They are priced per finished asset, and Runway's pricing page and docs disagree on the figures, so check the live rate in your portal.

Workflows are node graphs built in the Runway app and published as private, versioned endpoints you run with one POST, once your Runway workspace is linked to a Runway Dev project. Each node bills from Dev credits at its own price, graphs have no conditions or loops, and an invocation can report SUCCEEDED while single nodes failed, so check nodeErrors. Characters are real-time avatars built from one image on Runway's GWM-1 world model, with custom voices, knowledge documents and tool calling. Live sessions bill at $0.20 a minute and default to five minutes.

Shipping it: attribution, MCP and the right account

Any app built on the API must display a Powered by Runway credit with a link in its interface. Treat it as a launch requirement. Two MCP servers also confuse people. Runway MCP lets ChatGPT, Claude, Cursor and other assistants generate with your web-app account and its credits, with no API key. Runway Dev MCP serves coding agents instead: it researches models and prices, configures routers and Characters, and looks up failed tasks, but it generates nothing itself. The Runway Agent guide covers the in-app assistant route.

On self-serve plans, web-app credits and API credits are separate pools; Enterprise contracts can pool them and let admins move credits between workspaces and API projects. Enterprise on Runway Dev also adds a master services agreement, zero data retention support, uptime SLAs and volume pricing, which is sold only there. Runway's consumer Terms let it train on inputs and outputs, and a no-training commitment is an Enterprise term. Confirm which terms cover your project before sending client footage.

The bottom line

Runway Dev is a clean, prepaid way to put Runway's own models and a long list of other labs' video, image and audio models behind one key and one task format. The economics are easy to read: $0.01 a credit, a $10 minimum prepayment, and Gen-4.5 at $0.12 a second. The work is in the plumbing. Pin the version header, poll gently, download outputs within a day, treat moderation failures as spent money, and watch the tier caps before launch. For scripted video at volume it fits well; for occasional hand-made clips, a plan in the app is the simpler buy.