Summary: Runway’s Dev API changelog shows a steady fall-2026 cadence of new third-party models landing behind one API surface — Kling video, Ideogram image models, Nano Banana 2.1, Grok Imagine Video 1.5 Lite, and several others between late August and mid-October. For anyone building on generative media, the more interesting signal is the pricing and parameter detail published with every entry, which makes model swaps a spreadsheet exercise instead of a research project. Here’s what changed and how I’d read it if you’re picking models this quarter.
One endpoint, many vendors
The changelog at Runway Dev’s API Changelog & Updates documents a pattern rather than a single launch: roughly a dozen model additions since late August 2026, each with its own model identifier, parameter set, and per-second or per-image credit rate. Kling 3.0 arrived in three tiers (Standard, Pro, 4K) on October 13; Ideogram 4.5 and 4.0 landed the day before; Nano Banana 2.1 came October 6; GPT Image 2.5 in two variants on September 8.
The builder payoff is that the request shape stays consistent — text-to-video, image-to-video, and text-to-image endpoints — while the model slug and its provider-specific settings change. If you’ve already swapped one voice or video model behind a stable interface, the same discipline applies here. That’s the argument I made in One Voice API Swap, Same Price: a clean abstraction is what turns a vendor announcement into a one-line diff.
The pricing lines matter most
What separates a useful changelog from a marketing feed is the money column, and this one publishes it. A few figures worth noting from the supplied documentation:
- Kling 3.0 Standard: 9 credits per second without audio, 13 with; Pro is 12/17; 4K is 43 either way.
- Ideogram 4.5: quality tiers at 1, 4, 8, and 25 credits per image, multiplied by output count.
- Grok Imagine Video 1.5 Lite: 2 credits per second at 480p up to 14 at 1080p — with the caveat that 1080p is rendered at 720p and upscaled.
- Eleven v4 speech launched at an introductory 2.2 credits per 1,000 characters until October 12, 2026 PT, then 5.
That Eleven entry is a small case study in why you should read changelogs with a calendar open. The discount window expired one day after the Grok Imagine Lite entry shipped. If your ingestion is a weekly digest rather than a feed watch, introductory rates will pass you by.
Draft mode is the sleeper feature
The September 28 entry adds draft mode to Seedance 2.5, and it’s the change I’d prototype first. Pass draft: true for a 480p preview billed at the 480p rate; if it works, send the draftTaskId back alone to enhance to 1080p at the 1080p rate, reusing the original prompt, references, duration, seed, and audio setting. Previews stay enhanceable for 7 days.
For any pipeline where humans review generations before final render, this roughly halves the cost of the review loop. It’s the same economics as running cheap eval passes before expensive ones.
Pipeline-grade details, not just models
Not everything in the changelog is a new model. Ruby — Runway’s SDR-to-HDR converter — gained alpha channel passthrough on September 11 and, on September 12, plate-referenced ACEScg EXR sequences, after the earlier ACEScg outputs had lifted midtones in SDR views. The changelog explicitly documents what changed and why, which is rare and worth emulating in your own API notes.
Practical next step
Skim the changelog with three questions per entry: does the credit rate beat your current model at the same resolution, does the parameter set fit your existing request wrapper, and is there an introductory rate or a draft mode that changes the math. The supplied documentation is detailed enough to answer the first two without writing code — and if you can’t answer them, that’s a sign the model isn’t worth a spike yet.
Sources
View the original post on X