Runway Will License You Its Closed Weights. Read the Terms First.

Runway started selling its closed model weights to enterprises on September 11. On October 14, Baseten announced that customers can now run Gen-4.5 and Aleph 2.0 in single-tenant deployments, with fine-tuning on their own data. Runway posted the same news three minutes later, which tells you who was waiting on whom.
The details are easy to misread, so here's the structure. The licence runs from Runway to the customer. Baseten is the host, and Baseten's own access to those weights runs under a distribution agreement whose terms nobody has published. Runway is not licensing weights to Baseten in the sense that sentence usually implies. Fine-tuning happens under the customer's direct licence with Runway.
What Baseten adds is a channel. A customer who wants the weights now has a host that already runs GPUs and already knows how to do single-tenant deployments. That's a real obstacle removed. It's also not the interesting part.
What a buyer actually gets to read
The disclosed terms are thin. There's no price. There are no exclusivity terms. The one thing on record is the length: Runway's licensing page sells an annual licence with renewal built in. The 99.99% uptime SLA in the announcement is Baseten's commitment, not Runway's.
For a certain kind of buyer, that's fine. Banks, insurers, and defense-adjacent shops don't buy a model, they buy a contract with a named counterparty, an audit trail, and a place to send a lawyer. A hosted API can't be put in a data-residency filing. Weights in a single-tenant VPC can.
None of this is new for Runway. A September 11 post told enterprises they could license its model weights, fine-tune them, self-host, and sell what they make — 33 days before Baseten's announcement. The October news was a hosting channel, not a policy change. The press treated it as a launch because a launch is easier to write about than a procurement motion.
For everyone else, the missing number is the whole decision. You can't compare an annual licence with unpublished pricing against $0.40 a second on an API, or against a $0.06 image edit. You can only ask for a quote and see what comes back.

The ranking doesn't support the price
Here's where the story gets less flattering. On Artificial Analysis, read October 18, Gen-4.5 sits 26th at 1265 on the no-audio image-to-video board and 23rd at 913 on the no-audio text-to-video board. Aleph 2.0 is 8th at 1014 on the with-audio editing board, which is the one place Runway still looks like a leader.
So a company is licensing closed weights for a model ranked twenty-sixth, at a price it won't publish, alongside a promise that this is what serious enterprises do. That's not fraud. It's a different product category than a leaderboard position, and the two are being sold in the same conversation.
Runway's licensing page makes the argument directly: open weights hand you a weaker model and a to-do list. It's a good line, and it's true for a specific buyer. If you don't have a serving team, a quantisation pipeline, and someone who can chase a CUDA regression at 2am, an open checkpoint is a project, not a product.
The counterargument is that Runway is selling you someone else's to-do list, just with a support contract attached. Fine-tuning your own data on a 26th-place base model doesn't make the base model better than the models above it. It makes it yours, which is sometimes the only property that matters and sometimes the only property you paid for.
Runway is also a reseller now
The other half of the same week: Runway's API carries Kling models. Kling 3.0 and O3 at 4K run 43 credits a second, about $0.43. There's no Kling 4.0 row yet, which is consistent with Kuaishou still listing API access for that model as coming soon.
Read those two facts together and the company's position clarifies. Runway sells its own weights to enterprises, hosts competing models on its API, and runs a creative tool on top. That's not a model lab's shape. It's a distributor's shape, with a licensing business attached to the inventory it owns.
Distributors are durable. They also don't win on model quality, and they stop being the place people go first. The question for Runway is whether "the enterprise-grade place to run video models" is a big enough category to be worth more than "the best video model," and nobody has priced that yet.
Single-tenant is doing a lot of work in that sentence
That phrase is carrying more weight than it looks, because it's the only technical term in the announcement that changes what a customer can promise their own compliance team. Single-tenant means your deployment doesn't share a GPU pool or a request queue with anyone else. In practice that buys three things: predictable latency, no cross-customer logging, and a story you can tell an auditor about where the weights physically sit.
It does not buy you the ability to inspect the weights, retrain from scratch, or take the model elsewhere. A licence is a permission, not a title. If Runway's distribution agreement with Baseten changes, or if Runway decides a renewal price is higher than last year's, your fine-tuned adapter sits on someone else's inventory.
That's the asymmetry nobody has priced. An open-weight model can be forked when the vendor gets difficult. A licensed closed-weight model can only be renegotiated. Enterprises that have spent a decade learning this lesson with on-prem database vendors will recognize the shape of the trap, and the ones that sign anyway are effectively betting that video generation stays a procurement problem rather than a strategic one.
What to ask before you sign
Three questions, in order. What does the annual licence cost per GPU-hour of expected use, all in? What happens to your fine-tuned adapter if you stop renewing, and can you take it to another host? And which version of the base weights are you licensing, with what commitment to keep shipping improvements under the same terms?
If a vendor can answer all three in writing, the licence is a product. If it can only answer the first one verbally, you're buying a pilot with an invoice attached.
The broader shift is that "open versus closed" stopped being the useful axis. What matters now is what a contract lets you do with weights you can't audit. Runway is the first company at this quality level to put that on a price list, which makes it a test case rather than a trend. Whether the price list stays unpublished is the tell.
Related articles
The Video Leaderboard's Number One Is a Talking Head
HeyGen Video leads at 64,000 weekly requests, ahead of Veo and Wan. Chinese labs hold about 54.8% of routed video traffic while Runway sits at 0.3%. What request volume measures.
VideoGen Promises Ten-Minute Videos From One Prompt. Nobody Benchmarked It.
Ten minutes is assembled from many shots, not rendered in one pass. The product is the planning harness, and the cost and quality claims are still waiting on third-party validation.
NVIDIA's OpenShell Enforces Agent Policy in the Kernel
Kernel level checks on every file, syscall and connection, plus formal verification of policy changes before they apply. The install is one line. The operational bill is the rest of the page.
FLUX 3 Max Renders 10 Seconds of Video in 11. The Benchmark Has an Asterisk.
Fal post-trained FLUX 3 for speed and got a 1139 Elo claim out of it. The vote count is preliminary, the confidence interval overlaps the leader, and the 75% discount runs three weeks.