Runway

Runway AI, Inc. describes itself as "building foundational Real-World Intelligence that can understand, simulate and act in the world", and what a buyer actually purchases is access to generative media models. The same stack is sold three ways, and which door you pick changes the price, the controls and the contract.

The first is the web app and mobile apps, bought as Free, Standard, Pro and Max plans with a monthly credit allowance. The second is Runway Dev, a developer API with Node and Python SDKs, an OpenAPI reference and a go-live checklist, billed per use. The third is Runway MCP, a hosted MCP server that puts generation inside an assistant you already use.

Notably, the model list is not all Runway's own. The plan cards and the MCP page name Gen-4.5, Aleph 2.0 and Gen-4 Turbo beside Seedance 2.5, Kling 3.0, Veo 3.1, GPT Image 2 and Nano Banana Pro, and the API tab sells this as "Single vendor, unified billing across every model". Runway trains and serves its own models and resells others through the same meter.

The MCP server is real, documented and specific. It is listed in the ChatGPT app directory, and the FAQ names ChatGPT, Claude.ai, Claude Desktop, Grok Bot, Cursor and Replit as supported clients. Setup is three steps and the vendor is explicit that there is no developer work: asked whether an API key is required, the answer is "No. In ChatGPT, Grok Bot and Cursor, add the Runway plugin and sign in."

Two details matter more than the integration list. First, billing: "MCP uses your Runway credits, just like the Runway app", so an agent generating video is spending the same allowance as a person in the editor, with no separate budget or cap of its own. Second, entitlement. The marketing page offers MCP to anyone with "compatible agents", but the pricing table lists Runway MCP as a feature of Pro and Max only. It does not appear in the Standard or Free feature lists. A buyer reading the MCP page alone would reasonably expect a $15 plan to reach it.

The developer API has its own MCP guide for Cursor, Claude Code and Codex, which connects to a Runway Dev account rather than an app subscription. That is a separate path with separate billing.

The pricing page opens with its Monthly/Yearly control already set to Yearly, with the switch reporting checked and a "-20% off" label beside it. Read without operating the toggle, Standard looks like a $12 product. Clicking Monthly shows the real month-to-month rate of $15, and the annual card carries "Billed annually. Save $36/year." Pro is $35 monthly and Max is $95 monthly, saving $84 and $228 a year respectively when paid up front.

Credits are the unit, and their worth moves with the model. Standard's 625 monthly credits are quoted as "52s of Gen-4.5, 104s of Gen-4 Turbo, or 78 Gen-4 Images (1080p)". Max's 9,500 credits buy 791 seconds of Gen-4.5 or 475 Nano Banana Pro images at 2K. Only Max and Team roll unused credits forward, and only for one month.

The Free plan deserves a careful read. It is labelled "Free forever" at $0 a month, but the grant is "125 one-time credits", not a monthly allowance. It is a standing account rather than a trial, and also a one-shot sample. Team is $69 per seat monthly or $55 billed annually, starts at two users and pools 165,600 credits a year across the workspace. Enterprise is quote-only.

The API is the one genuinely pay-as-you-go path: video from $0.05 per second, images from $0.01 and audio from $0.0025 per second, free to start and with no credit card asked for.

Section 4.4 of the Terms of Use is the clause to read before uploading client material. Runway does not claim ownership of inputs or outputs, and states it "does not restrict your commercial use of your Outputs", which is unusually clear and applies from the Free plan up. The same clause then grants the company an "irrevocable, perpetual, worldwide, royalty-free" licence to use those inputs and outputs to "train and improve its AI models".

There is no settings toggle offered against this. "No training on your data" appears once on the site, as a row in the Enterprise column of the Team and Enterprise comparison, alongside "Support for zero-data retention" in the API tab's Enterprise card. For everyone on a self-serve plan, the published position is that submitted material may be used for model training.

The data security page states a "Technical separation between customer content storage and model training systems" and that "Customer content is never used for testing new features or changes". Neither sentence contradicts section 4.4; they describe isolation and test practice, not a training exclusion.

Security documentation here is better than most listings of this size. Runway publishes a dated security page, last updated August 2026, and names SOC 2 Type II with annual recertification and ISO/IEC 27001 issued by an accredited body, both requestable through a trust centre. It documents TLS 1.2+ in transit, AES-256 at rest, role-based access with enforced MFA, a private bug bounty programme, external penetration testing, a published subprocessor list and a public status page.

What is absent is the paperwork a procurement team asks for next. GDPR appears as "GDPR Readiness" and CCPA as "CCPA/CPRA Controls", which are statements of alignment rather than an offer to sign a data processing agreement. No DPA is published or named anywhere on the site. Legal indemnification and "Enterprise terms & MSA" are Team and Enterprise rows.

Data location is fixed. The privacy policy states that Runway is headquartered in the United States, may use vendors in other countries, and that using the service constitutes consent to transfer "in and to the United States and other applicable countries in which the privacy laws may not be as comprehensive". No region can be chosen. Retention is set by the vendor "for only as long as necessary", with face scans and voiceprints held no longer than three years from last interaction.

A team that needs its material kept out of model training cannot get that on a self-serve plan at any price. Standard, Pro and Max all sit under the same section 4.4 licence, and the exclusion is an Enterprise negotiation. The same is true of an uptime commitment, zero-data retention, single sign-on and admin roles: each is a Team or Enterprise row, so a small studio with compliance obligations is pushed into a sales conversation rather than a plan.

Anyone budgeting tightly should model the credit maths before subscribing. Because a credit buys different amounts of output depending on model and resolution, and because MCP and the editor draw on the same pool, a month of experimentation can exhaust an allowance with no published calculator to predict it. Top-ups are available from Pro upward, which makes overspend easy and forecasting harder.

Buyers expecting the MCP server at the entry price should check the plan table rather than the MCP page. And a team hoping to self-host, or to inspect the models, will find neither: the Terms forbid reverse engineering the service, and nothing on the site discusses customer-run deployment in either direction.