The Production Studio That Also Hosts Its Rivals
Runway trains on your work unless you hold an Enterprise contract. Section 4.4 of its Terms of Use, last updated September 15, 2026, says inputs and outputs may be used to train its models. No opt-out. No paid-tier carve-out. Free, Standard, Pro, Max and Team all sit inside it. Price that in first. Then enjoy the range, because it’s real: Gen-4.5, Aleph 2.0, Agent, and rival models from Google, ByteDance, Kuaishou, MiniMax, xAI and Alibaba in one workspace. Since October 13, 2026, seven Kling models are callable from the REST API too. And Unlimited moves to Max on November 30, 2026, 46 days after October 15.
A studio, an aggregator and a marketing agent in one
If your mental model is “Runway is where you make Gen-4 clips,” it’s out of date. In 2026 Runway is three things at once. A model maker. A shop that stocks other vendors’ best models. And an agent that produces finished campaigns. So choosing Runway isn’t choosing a generator. It’s choosing a pipeline, right? Judge it like one.
Control-first and production-leaning
Runway rewards direction. References, camera paths, motion brushes and edit-grade tools that slot into a real pipeline instead of sitting beside it. Since late August it also finishes in professional formats, with HDR output and ACEScg OpenEXR delivery. One catch. Those finishing formats are a plan entitlement in the app, not a universal feature. Runway’s pricing page names HDR, ProRes and image sequences as studio-grade features of Max, and the product changelog tags the Ruby converter for Max and above.
- Repeatable shots, not just vibes
- Character and location consistency
- Six access surfaces, different model lists
Finished, on-brand video
Campaign assets. Product and brand video. Social cut-downs, avatars and explainer beats that have to survive a review and still ship.
- Agencies and in-house brand teams
- Repeatable campaigns, not one-offs
- Generation through final export
You need the pipeline, not one clip
Reach for Runway when you want generation, a deep bench of third-party models, editing and finishing under one login. And when brand consistency matters more than novelty. One condition. Runway trains on self-serve accounts, so work you cannot let into a training set belongs on an Enterprise contract.
- Many models, one workspace
- Edit and finish in the same tool
- Enterprise contract if training is a problem
Kling on the API, pooled Enterprise credits, and a deadline that held
Kling reached the REST API on October 13, 2026. Runway’s API changelog adds seven identifiers, now on the models and pricing pages. Three older facts set up everything else. Section 4.4 of the Terms of Use, updated September 15, 2026, puts every self-serve user’s inputs and outputs into training. The identifiers gen4_aleph and gen3a_turbo were declared no longer available on July 30, 2026. And Unlimited moves to Max on November 30, 2026, with the annual refund only before it. The rest of the news between September 4 and October 8, 2026 is commercial: a Team plan, two editor plugins, Enterprise credits that pool across app and API, and new partner routes through OpenAI and Anthropic.
Runway trains on self-serve inputs and outputs Terms of Use, September 15, 2026
Read this before anything else about Runway. Section 4.4 of the Terms of Use says inputs and outputs may be used to train and improve Runway’s models, with labeling and content moderation named too. No opt-out switch. No paid-tier carve-out. Free, Standard, Pro, Max and Team all sit inside that sentence. The no-training commitment is a contract line instead of a plan feature. Runway’s Enterprise Services Terms say Runway may not use Customer Content as training data for the Services. So the question isn’t which plan you buy. It’s whether you’re on a contract at all. If client footage, unreleased product or a face under NDA goes into Runway on a credit card, assume it can be trained on.
· Source: Runway Terms of Use, Section 4.4, last updated September 15, 2026, and Runway Enterprise Services Terms, last updated June 1, 2026
· Section 4.4, verbatim: “You acknowledge that Inputs and Outputs may be used by the Company to train and improve its AI models, algorithms and related technology, products and services (including for labeling, classification, content moderation and model training purposes).”
- Every self-serve plan is trained on
- No opt-out published on any first-party page
- The exemption lives in the Enterprise contract
Kling reaches the REST API October 13, 2026
Runway’s API changelog entry of October 13, 2026 says Kling is now available on Runway Dev. Seven identifiers, all on the API models and pricing pages: kling2.5_turbo_pro at 12 credits per second; kling3.0_standard at 9 without audio and 13 with; kling3.0_pro at 12 and 17; kling3.0_4k at 43; klingO3_standard at 13, klingO3_pro at 17 and klingO3_4k at 43. Watch the audio default. It’s on, and Kling 3.0 Standard and Pro only bill the lower rate when audio is set to false. So a developer who forgets the flag pays over 40 percent more per second. Not every Kling made the trip. Kling 3.0 Turbo, Kling 3.0 Motion Control, Kling 2.6 Pro and Kling 2.5 Turbo still aren’t on the REST API. And no Sora identifier appears on the API pages or Academy.
- Seven Kling identifiers on Runway Dev
- Audio defaults on, so turn it off to pay less
- Turbo, Motion Control, 2.6 Pro and 2.5 Turbo stay off the API
Enterprise credits now pool across app and API September 18, 2026
For a self-serve account, the two kinds of Runway credit never mix. Runway’s developer docs say web app credits and API credits are separate pools. Its help center says web app credits will never appear in your API credits, and vice versa. Enterprise changed on September 18, 2026. Runway’s Unified Pricing post says Enterprise credit commitments are now fungible across the web app and Runway Dev, with one invoice and a workspace analytics view where admins move credits between web app and Dev workspaces. Two dated offers ride on it. Sign by December 31, 2026 for 10 percent more credits on signing. And an existing Enterprise customer whose referral leads to a purchase by December 31, 2026 gets 5,000 complimentary credits and one free Applied AI Architect day. If your team runs a creative seat and an integration, this is what stops you budgeting twice.
- Self-serve: two separate pools
- Enterprise: one pool since September 18, 2026
- 10 percent more credits for signing by December 31, 2026
September 1 came and went Help article, updated August 27, 2026
The help article “Unlimited plan is switching to Max” was edited on August 27, 2026. That’s five days before the September 1 deadline it had carried until then, and September 1 no longer appears anywhere in it. Read October 15, 2026, the November 30 date hasn’t moved again. The article says three things. On November 30th Unlimited subscribers move to Max. Subscribers move automatically at the same monthly price. And Runway is extending Unlimited access for eligible legacy Unlimited subscriptions through November 30th, 2026. The tense is still future, so nothing has migrated yet. And the article carries an update stamp rather than a publication date. That’s how a date can change without leaving a trace, right? Open it yourself before you act on the date.
- New date: November 30, 2026
- No announcement and no changelog entry
- Legacy Unlimited subscriptions stay active until then
Gen-4 Aleph is gone, not going Executed July 30, 2026
This is the removal with the most direct cost to a reader, and the one Runway gave no warning of. Anyone who read gen4_aleph as merely deprecated and waited for a shutoff notice was waiting for a notice that never came. The shutoff had already happened. Runway’s API changelog entry of July 30, 2026 records that Gen-3 Alpha Turbo (gen3a_turbo) and Gen-4 Aleph (gen4_aleph) are no longer available. Read October 15, 2026, neither identifier appears on the API models page or the API pricing page. Both of those pages are rewritten where they stand, so confirm directly before you schedule anything. And if a job, script or scheduled render still names either model, it isn’t deprecated. It’s failing.
- Both identifiers removed on the same day
aleph2is the replacement for Aleph work- Audit running jobs now, not at the next release
Gen-4.5 Flagship
Runway’s best video model for text- and image-to-video. Introduced December 1, 2025 and still the top of the line. There’s no Gen-5. Runway’s own wording is narrower than the summary that gets repeated: it calls Gen-4.5 the world’s top-rated AI video generation model, ranked number one on the Artificial Analysis Text-to-Video benchmark with 1,247 Elo points. That’s a leaderboard position on one benchmark, so test it on your own shots. Since August 20, 2026 it can deliver straight into HDR and post-production formats. On the API those formats cost extra. ProRes and PNG sequences add 5 credits per second. 10-bit and HDR add 20, or 40 above 4 megapixels. So a 4K HDR second of Gen-4.5 is 52 credits, not 12.
- Text or image to video
- Finishing formats add a per-second surcharge
- Vendor claim, so test on your shots
Aleph 2.0 and Edit Studio Editor
Change one frame and Aleph 2.0 changes the rest of your video to match, up to 30 seconds. Edit Studio shows the edit as an image first, so you can shape the look before you commit to generating. The API model is aleph2, and it joined Gen-4.5 in the August 20 HDR output update.
- Edit-grade video-to-video
- Preview before you generate
gen4_alephwas switched off on July 30, 2026
Agent 2.0 Campaign engine
Runway’s product page calls Agent a chat-based marketing agent that builds and scales the work: concepts, ads, full campaigns and variants. Agent Skills shipped in July 2026 to run those jobs from one command. Projects and Brand Kits came to Agent on August 11, 2026, tagged for Enterprise, and Agent Brand Kits reached all plans on September 10. Here’s the catch. The Agent product page is undated and records none of that, so the changelog is the only place Agent’s development shows up at all.
- Brief to finished campaign
- Ad campaigns, commercials, localization
- Agent Brand Kits on all plans since September 10, 2026
Studio Finishing
A built-in timeline to trim, stitch, reorder and export a final video without leaving Runway. Generation and finishing now live in one tool.
- Real editing timeline
- One-tool round trip
- Export production-ready cuts
Characters and GWM-1 Real-time
Launched in March 2026. Real-time intelligent avatars you can talk with, running on Runway’s GWM-1 general world model. Runway quotes roughly 1.75s response, 24fps and 37ms per frame. Those are vendor figures, not independent tests.
- Conversational avatars
- API model gwm1_avatars
- Consent and likeness rules apply
Veo 3: sunset, executed August 4, 2026
Runway’s API model guide said Veo 3 was deprecated and would be officially sunset on August 4th, 2026. It was. Read October 15, 2026, veo3 appears on neither the API models page nor the API pricing page. The same holds for the July 30 removals, though those were recorded on the day rather than published ahead of it. So Veo 3 is the one removal Runway dated in advance and then executed on that date. That’s the instructive contrast with the Unlimited migration. A technical sunset published in the docs held. A commercial date published in a help article moved. The Veo 3.1 pair, veo3.1 and veo3.1_fast, is unaffected and remain the upgrade path, and veo3.1_fast is a lot cheaper at 15 credits per second with audio and 10 without.
veo3calls hit a model that’s goneveo3.1_fastis the cheap half of the pair- A sunset date published in advance held; not every removal gets one
Hosted third-party models Rented
One subscription, a lot of outside models. Google Veo 3.1 and Veo 3.1 Fast. ByteDance Seedance 2.5 and Seedream 5.0. Kling 3.0 and Kling O3. MiniMax H3 and, since September 3, H3 Max. xAI Grok Imagine video and image. FLUX.2, GPT Image 2 and the September 8 GPT Image 2.5 pair. Ideogram. ElevenLabs audio, with ElevenLabs V4 since September 29. And since August 26, Alibaba WAN 3.0. Two traps in that list. Muse Image is an API-side model. It’s priced on the API pricing page at 1 credit per image and doesn’t appear on Academy at all, so it’s easy to file on the wrong side of the app-versus-API split. And the Seedance 2.5 identifier, published on the API pages rather than in the app, is seedance2_5. The roster moves in both directions. Sora 2 and Sora 2 Pro went on April 3, 2026, and no Sora identifier appears on the API pages or Academy. Veo 3 went on August 4. And Kling went the other way, reaching the REST API on October 13.
- Best model per shot
- Muse Image is API-only, not app-side
- Models can be pulled, so name a fallback
What arrived in late August API changelog
Five dated additions inside eleven days. August 20 brought ruby, which Runway describes as converting any SDR video, whether uploaded or generated by any model, into true HDR (paraphrase). Its price isn’t a range. It’s a two-step conditional. The API pricing page sets it at 20 credits per second, rising to 40 credits per second when the source is larger than 4 megapixels, which is roughly 4K. So budget 40 for a 4K plate. The same day brought HDR output formats for Gen-4.5 and Aleph 2.0. August 26 added WAN 3.0 as wan3 and Muse Image as muse_image. August 31 added ACEScg OpenEXR delivery, in ACES 1.3 and ACES 2.0 variants. And Runway’s product changelog tags the Ruby entry Max and above, which makes it a plan entitlement in the app as well as a line item. Pro doesn’t reach it there.
ruby: 20 credits per second, 40 above 4 megapixelswan3: up to 30 seconds with native audio- Ruby is tagged Max and above in the app
Solaris is research, not a product August 31, 2026
Runway published a research post, “Introducing Solaris,” on an interface world model. Research only. No product, no API model, no pricing, and it’s explicitly not a Gen-5 announcement. Read it as a direction of travel, not as something you can put into a workflow this year.
- Research post, August 31, 2026
- Not a product and not Gen-5
- Nothing to integrate yet
GWM Worlds 2 September 3, 2026
Three days later came another research post, and again not a product. Runway describes GWM Worlds 2 as extending the GWM Worlds work shown in December, now with generated audio and richer subject and scene control. It delivers interactive worlds generated in real time as continuous 720p video at 24 frames per second, with audio at 48,000 Hz that responds to your inputs as you explore. Those are vendor figures. There’s no API model, no pricing and, again, no Gen-5.
- Research, not a product
- Real-time interactive worlds with audio
- Nothing to procure and nothing to integrate
The product changelog is the buyer’s changelog Undated page
Runway runs two changelogs, and the one at runway.com/changelog is the one a buyer needs. Read October 15, 2026, its newest entry is dated October 2: Seedance 2.5 Draft Mode in Agent, and Ideogram 4.5 in Runway MCP. It beats the API changelog for a buyer because every entry carries a tag: All Plans, Paid Plans, Pro+ Plans, Max+ Plans, Enterprise or API. That’s where the HDR, ProRes, Ruby and Enhance Frame Rate gating described here comes from, and it’s why Ruby lands on Max rather than Pro in the app. One warning. The page carries no history, and its count of entries per month doesn’t hold still between readings. So you can’t tell from the page alone what was added late or taken down. Screenshot the entry you plan to rely on.
- Two changelogs, different scopes
- Tags include Pro+ Plans, Max+ Plans and API
- No history, so screenshot what you cite
The Team plan September 4, 2026
A self-serve tier between Pro and Enterprise, and the biggest pricing change of the summer for a small studio. Runway prices Team at $69 per seat monthly, or $55 per seat on annual billing, for 2 to 9 seats. Every seat adds 6,900 credits a month into one shared pool, and unspent credits carry over for one month. You also get 1TB of storage, up to 100 shared projects, shared Brand Kits, up to 20 agent skills, and Agent Connectors for Figma, Dropbox and Notion. The catch isn’t the price. It’s discovery. On runway.com/pricing, read October 15, 2026, the default view is the Individual table and Team sits behind a Team & Enterprise tab, so a buyer comparing tiers can miss the tier that fits them. Every Team figure here comes from the September 4 announcement, which is the dated record.
- $69 per seat monthly, $55 annual
- 6,900 credits per seat into a shared pool
- Hidden behind a tab on the pricing page
Runway for Premiere Pro and After Effects September 8, 2026
An access surface inside the editor, beside the web app, the REST API and the two MCP servers. Runway describes generating images and video, restyling the footage on your timeline, and dropping every result straight into your comp or sequence. Aleph 2.0 is the restyle path: select a clip, restyle its anchor frames, and Aleph 2.0 re-renders the full clip to match. The installer is free for macOS and Windows. Runway’s help article says generating needs a Standard plan or higher and spends your plan credits. The announcement names only Aleph 2.0, but runway.com/plugins, read October 15, 2026, lists nine models. Video: Gen-4.5, Seedance 2.5, Kling 3.0 Pro and Veo 3.1. Stills: Gen-4 Image, Nano Banana and GPT Image 2. Restyle and HDR: Aleph 2.0 and Runway Ruby. That last one is the catch. The app tags Ruby Max and above, and neither plugin page says whether that gate applies inside the panel.
- Free installer, Standard plan or higher to generate
- Nine models listed, Kling 3.0 Pro among them
- Ruby is listed; its plan gate in the panel isn’t stated
Runway for DaVinci Resolve September 23, 2026
A second editor door, two weeks after the first. Runway describes generating images and video, restyling the clip under your playhead, and importing every result straight into your Media Pool and onto your timeline. Aleph 2.0 is the restyle path again, with more control than the Adobe version advertises: restyle the anchor frames, or set up to five keyframes, and Aleph 2.0 re-renders the clip to match at the same duration as your source. The installer is free for macOS and Windows and needs DaVinci Resolve Studio 19 or later. Generating is available on paid Runway plans and uses your existing plan credits. Runway’s plugins page names Resolve as a host of the same panel and its nine-model list, and Runway’s Resolve help article says the plugin carries the same lineup as your plan on web. Neither page says every listed model works inside Resolve, so open the panel before you promise one.
- Free installer, Resolve Studio 19 or later
- Up to five keyframes on an Aleph 2.0 restyle
- Shares the plugin model list on runway.com/plugins
Runway inside other vendors’ products September 29 to October 8, 2026
Three new routes in, and none of them is a Runway door in the old sense. September 29: Runway joined the OpenAI Marketplace, so an eligible enterprise can apply part of its OpenAI commitment toward Runway purchases. October 1: the product changelog records Runway inside OpenAI’s Dots, for overnight plan-and-shoot ad production. October 8: Anthropic launched Claude Motion with Runway as a launch partner, and any animation made in Claude Motion can be exported to Runway. Claude Motion is in beta on Claude Team and Enterprise plans. The procurement point is the first one. If your company already holds an OpenAI commitment, ask whether Runway spend can draw it down before you sign a separate contract.
- OpenAI Marketplace, September 29, 2026
- Claude Motion exports to Runway, October 8, 2026
- An OpenAI commitment may pay for Runway
Runway Dev MCP September 2, 2026
A second MCP server, and not the same product as the consumer Runway MCP. Runway describes it as a hosted MCP server that connects your Runway Dev account to the coding tools you already work in, naming Claude, ChatGPT, Codex and Cursor. From inside the editor an agent can research which models a project can use along with prices and capabilities, configure model routing, and debug an integration. So the developer surface now has an agent front door of its own. If a team is scoping Runway from an IDE, this is what answers “which models can this project actually call” without a human reading two documentation pages.
- Separate from the consumer Runway MCP
- Claude, ChatGPT, Codex and Cursor
- Model research, routing and integration debugging
What the API costs, and how fast you can call it Developer docs
Runway publishes an API credit price in its developer docs that its plan table never mentions. Credits can be purchased for $0.01 per credit in the developer portal for a project, and a minimum payment of $10 is required to get started. So a 10-second Gen-4.5 generation is 120 credits, which is $1.20. Then throughput, which is where most integration plans go wrong. There are five spend tiers. Tier 1 allows 1 to 2 concurrent generations against a $100 monthly spend cap. Tier 5 allows 20 against $100,000. Advancement is automatic, immediately after a project has purchased $50, $100, $1,000 and $5,000, with no waiting period. The part worth repeating to an engineer: there is no maximum requests-per-minute limit, as long as requests stay inside the maximum daily generations limit. Those daily caps run from 50 to 200 at Tier 1 up to 25,000 to 30,000 at Tier 5, depending on the model class.
· Source: Runway Dev pricing, setup and tiers pages, rolling and undated, read October 15, 2026
- $0.01 per credit, $10 minimum to start
- Concurrency 1 to 2 at Tier 1, 20 at Tier 5
- No requests-per-minute cap, only a daily generation cap
Enterprise: indemnity, no training, and where each promise actually lives August 20, 2026
The most useful Runway post of 2026 for a procurement team, and the easiest to miss, because it sits on Runway News rather than on the pricing or enterprise pages. It commits to uncapped IP indemnification with third-party models included, to no training on your data, and to security standards it lists as SOC2, ISO 27001, GDPR and CCPA, zero data retention and SSO. Every one of those is an Enterprise commitment, and the no-training line reads against Section 4.4 of the general Terms of Use, which says the opposite for everyone else. The governing document is the Enterprise Services Terms, last updated June 1, 2026. It says Runway may not use Customer Content as training data for the Services, and that as between the parties the customer owns all rights, title and interest in Customer Content. Section 9.2 caps each party’s total liability at what the customer paid Runway in the preceding 12 months, then carves out both parties’ IP indemnification obligations along with willful misconduct and gross negligence. That carve-out is what makes the indemnity uncapped, and it’s checkable rather than a marketing adjective. One mismatch to expect in a review. The enterprise page at runway.com/enterprise, read October 15, 2026, names SOC 2 Type 2, no training on your data, full ownership of your outputs, customizable access and permissions, and SSO. ISO 27001, GDPR, CCPA and zero data retention show up in the August 20 post and not on that page, so ask for each one by name.
· Source: The Next Phase of Enterprise Video Generation, Runway News, August 20, 2026; Runway Enterprise Services Terms, last updated June 1, 2026; runway.com/enterprise, rolling and undated, read October 15, 2026
- Uncapped IP indemnity rests on the Section 9.2 carve-out
- No training is contractual, not a paid-plan feature
- The enterprise page lists fewer standards than the post
Workspace and admin features only the changelog records August and September 2026
Runway’s product changelog, read October 15, 2026, carries a run of workspace and governance items Runway announced nowhere else. Model Router Routing History on August 3. Comments on Assets on August 7, tagged Enterprise. Projects and Brand Kits in Agent on August 11, tagged Enterprise. Compositing Nodes on August 18. On August 24, both Default SSO for New Workspaces and Org Config Alerts, Enterprise again. And Agent in Asset Comments on September 15, tagged Enterprise. None of them changes a model or a price, which is exactly why they stay invisible outside the changelog. All of them are the kind of thing an administrator gets asked about in a security review. The changelog is rewritten in place, so confirm the entries and their plan tags yourself.
- Default SSO for new workspaces, August 24
- Org Config Alerts for administrators, August 24
- Comments on Assets and routing history for auditability
Model licensing changes the shape of the deal August 20, 2026
The same post announces something in a different category from every other way of buying Runway. Under a model licensing program, Runway delivers closed model weights and a proprietary training and inference framework straight to a customer. That customer hosts the model inside their own environment and fine-tunes it on their own data. Every other commercial shape is a rental. A model you reach over an API or inside an app, on terms Runway can change, on a roster Runway can shorten. This one isn’t. So if your objection to Runway has been that a hosted model can be withdrawn, this is the published answer. It’s an enterprise conversation, not a plan you can buy on the pricing page.
- Closed weights delivered to the customer
- Hosted in your environment, fine-tuned on your data
- Not a self-serve tier, so pricing is unpublished
What arrived in September API changelog
Five dated API items between September 3 and September 17, read October 15, 2026. September 3: MiniMax H3 Max as h3_max, 5 to 15 second durations at 5 credits per second for 480p and 8 for 768p, and a distinct identifier from hailuo3, which is still live. September 8: GPT Image 2.5 as gpt_image_2_5_flare and gpt_image_2_5_sunburst. Up to 16 reference images, 30 aspect ratios and 1 to 10 outputs per call, billed 1 to 76 credits per image plus 1 credit per reference. September 11: Ruby now preserves a source video’s alpha channel on the video-to-HDR route, writing alpha as the EXR A channel and using ProRes 4444. September 12: both ACEScg EXR sequence outputs on the Ruby endpoint, hdr_exr_acescg_sequence_1_3 and hdr_exr_acescg_sequence_2_0, are now built from the source plate, with colorimetry metadata updated to match. September 17: a new model, enhance_frame_rate, retiming a video to any of ten target frame rates at 1 credit per 2 seconds, with inputs of at most 300 seconds.
h3_maxisn’thailuo3- Ruby gained alpha and plate-referenced ACEScg in two ACES versions
enhance_frame_rate: ten targets, including drop-frame
Runway-native models, the rivals it hosts, and an API with a different list
Six doors that Runway runs itself, and the menus still differ. The web app. The REST API. Runway MCP. Runway Dev MCP, added September 2, 2026. The Adobe plugin from September 8. And the DaVinci Resolve plugin from September 23. Partner routes through OpenAI and Anthropic sit outside them. Kling shows how fast a menu changes. On October 13, 2026 seven Kling models reached the REST API, while Kling 3.0 Turbo, Motion Control, 2.6 Pro and 2.5 Turbo stayed off it. Meta’s muse_image runs the other way, API-only. And Academy still prices Gen-4 Aleph, which the API retired on July 30, 2026. Picking the model per shot is the craft. Knowing which door it sits behind is the procurement question.
Gen-4.5
The highest-quality Runway video. Reach for it when the shot ships, now with HDR and post-production output formats.
- Text or image to video
- First-frame conditioning
Gen-4 Turbo
Faster, cheaper Runway video for iteration and look-finding before you commit credits.
- Image to video
- Test cheap, then upgrade
Act-Two
Motion-capture performance with head, face, body and hand tracking. Drive a character from a reference performance.
- Performance capture
- Consent required
Gen-4 Image and References
Runway’s image model with reference-driven characters and locations, plus Layout Sketch for composition control.
- Consistent characters
- Draw to compose
Aleph 2.0
Prompt-driven editing of footage you already have, up to 30 seconds, with an image preview in Edit Studio before you generate.
- API model aleph2
- Preview then commit
Ruby
Added August 20, 2026. A conversion model, not a generator. It takes SDR video, uploaded or generated by any model, and produces HDR (paraphrase of Runway’s wording). On the API it takes SDR inputs of 30 seconds or less, up to 4K, through /v1/video_to_hdr. Since September 11 it preserves a source video’s alpha channel, writing alpha as the EXR A channel and using ProRes 4444. The API pricing page, read October 15, 2026, sets the rate at 20 credits per second, and 40 credits per second when the source is larger than 4 megapixels, which is roughly 4K. The models page now describes the output in one line: HEVC, ProRes or an EXR sequence. The identifiers live on the Professional outputs page and in the API changelog, which names both hdr_exr_acescg_sequence_1_3 and hdr_exr_acescg_sequence_2_0 on the Ruby endpoint. In the app, the product changelog tags Ruby Max and above. On Runway Dev, no page states a plan requirement.
- API model
ruby - 20 credits per second, 40 above 4 megapixels
- ACES 1.3 and ACES 2.0 outputs
Enhance Frame Rate
Added September 17, 2026. A second conversion model: it retimes a video to a target frame rate. On the API there are ten targets, and the three a broadcast deliverable usually needs are the drop-frame ones. The full set is 24, 25, 30, 48, 50, 60, 120, 23_98, 29_97 and 59_94, with inputs of at most 300 seconds, per the September 17 API changelog entry. The models page now gives it a single line. Priced at 1 credit per 2 seconds, and Runway’s launch post says that holds regardless of resolution or target frame rate. The app side differs twice. The product changelog lists 25, 30, 48, 60, 120 and 59.94 fps, up to 4K and 5 minutes, and tags the entry Max+ Plans. So the door decides both the targets you get and the plan you need.
- API model
enhance_frame_rate - 1 credit per 2 seconds, inputs up to 300 seconds
- Tagged Max+ Plans in the app
The rest of the native bench
These are the native identifiers easiest to overlook, because the app never names them. All of them were on the API models page, read October 15, 2026: act_two, gen4_image, gen4_image_turbo, ruby and enhance_frame_rate. If you’re scoping an integration, start from this list rather than from the app.
- Native IDs, not third-party ones
gen4_image_turbois the cheap image seat
Characters and GWM-1
Real-time conversational avatars built on Runway’s general world model, available on the web and through the API.
- API model gwm1_avatars
- Rights and consent first
| Category | Hosted in the app, with Academy credits per second unless noted | Reach for it when |
|---|---|---|
| Video: Kling O3 and 3.0 | Kling O3 (4K 43, Pro 17, Standard 13); Kling 3.0 Pro (12, or 17 with audio); Kling 3.0 Std (9, or 13 with audio) | Timed keyframes or multi-shot with audio; on the REST API since October 13, 2026 |
| Video: Kling Turbo, Motion Control and 2.x | Kling 3.0 Turbo (720p 14, 1080p 16); Kling 3.0 Motion Control (720p 13, 1080p 17); Kling 2.6 Pro (25); Kling 2.5 Turbo Pro (720p 12, 1080p 15); Kling 2.5 Turbo (10) | Motion control or an earlier Kling look; only 2.5 Turbo Pro is on the REST API |
| Video: Google and ByteDance | Google Veo 3.1; ByteDance Seedance 2.5 (480p 20, 720p 30, 1080p 68, plus 10, 15 or 34 per second of input video); Seedance 2.0, 2.0 Fast and 2.0 Mini | Motion, audio or a look Gen-4.5 does not deliver |
| Video: Alibaba and MiniMax | Alibaba WAN 3.0 (480p 5, 720p 10, 1080p 20), WAN 2.6 and WAN 2.6 Flash; Alibaba HappyHorse 1.0; MiniMax H3 (768p 10, 2K 15; matches hailuo3, which adds 2 per reference on the API) | Low per-second rates where resolution is the cost lever |
| Video: Lightricks, Black Forest Labs, xAI | Lightricks LTX 2.5 Fast (720p 9, 1080p 13, 1440p 19, 4K 30) and LTX 2.5 Pro (720p 12, 1080p 17); Black Forest Labs FLUX 3 (17); xAI Grok Imagine | A published 1440p or 4K video rate |
| Video: retired from the API | Gen-4 Aleph (15; the API retired it on July 30, 2026) | Not for new work; aleph2 replaces it |
| Image: ByteDance and Google | ByteDance Seedream 5.0 Pro and Lite; Seedream 5.0 Pro Layers (18 per run); Characters (2 per 6 seconds); Google Nano Banana, Nano Banana Pro, Nano Banana Pro 2 and Pro 2 Lite (Runway’s labels, not Google’s) | Stills, first frames or style references |
| Image: OpenAI and Ideogram | OpenAI GPT Image 2.5 Flare and Sunburst (1K and 2K 16, 4K 19); GPT Image 2, 1.5 and 1 Mini; Ideogram 4 (Turbo 4, Default 7, Quality 11); Ideogram P-Image | Stills, first frames or style references |
| Image: Black Forest Labs, xAI, Magnific | FLUX.2 Max and Klein; xAI Grok Image Quality and Grok Imagine Image 2.0; Magnific Style Transfer | Stills, first frames or style references |
| Audio | ByteDance Seed Audio 1.0; Google Lyria 3 Pro and Lyria 3 Clip; ElevenLabs text-to-speech, sound effects, dubbing, isolation and speech-to-speech, plus Eleven 3v; Fish Audio (1 credit per 250 characters) | Voice, music, sound design, dubbing or cleanup |
| Upscale | Topaz 4K; Magnific Precision | A finished cut has to hit delivery resolution |
| Language | Google Gemini 2.5 Flash; Anthropic Claude Sonnet 4.5, Opus 4.5 and Opus 4.8 | Scripting, briefs and copy inside the same workspace |
Veo negative prompts
The API added an optional negativePrompt field on the Veo models, for both text-to-video and image-to-video, up to 1,000 characters. Useful for killing a recurring artifact without rewriting the whole brief. Since the August 4, 2026 sunset it applies to the veo3.1 pair only.
- Text-to-video and image-to-video
- Up to 1,000 characters
veo3.1andveo3.1_fast
Seedream 5.0 Lite in the API
Seedream 5.0 Lite joined the API as seedream5_lite: prompts up to 4,000 characters, up to 14 reference images, 4 credits per image. The cheap end of the image bench is now reachable from code, not just from the app.
- 4,000-character prompts
- Up to 14 reference images
- 4 credits per image
Gen-4 Aleph is gone, with nothing published ahead of it
Runway’s API changelog entry of July 30, 2026 states that Gen-3 Alpha Turbo (gen3a_turbo) and Gen-4 Aleph (gen4_aleph) are no longer available. Neither identifier is on the API models page or the API pricing page, read October 15, 2026. No deprecation window came first. The changelog recorded the removal on the day it took effect, so anyone treating gen4_aleph as deprecated and waiting for a shutoff notice was waiting for something Runway never published. Aleph work moves to aleph2. Gen-3 Alpha Turbo has no like-for-like successor, so re-scope rather than re-point. Beyond that, no new deprecation, sunset or removal was published on either API page or in any API changelog entry between August 19 and October 15, 2026. The changelog does record aleph2_alpha as a deprecated alias, dated June 2, 2026, so check any code that still calls it. Its newest entry is the October 13 Kling addition.
- Executed July 30, 2026, not pending
- Move Aleph work to
aleph2 - No new deprecations published through October 15
The one model with a date on it kept the date
Runway’s API model guide gave Veo 3 an August 4, 2026 sunset and recommended Veo 3.1 or Model Routers. The model is still absent from that guide and from API pricing. Worth recording as evidence rather than trivia: the published technical date was accurate to the day, so a team that keeps a deprecation watchlist migrated calmly.
- The published technical date held
veo3.1is the upgrade- Model Routers are Runway’s other suggestion
Four models, a resolution addition, and two changelogs that disagree
The API changelog surfaced Seedance 2.5 as seedance2_5 and Grok Imagine Video 1.5 on August 7, and Seedance 2.5 at 1080p on August 15, where Runway notes that 4K stays on Seedance 2.0. MiniMax H3 arrived as hailuo3 and Grok Imagine Image 2 as grok_imagine_image_2. Now the part worth a line of its own. Runway’s two changelogs date the same arrivals differently. The product changelog lists MiniMax Hailuo 3.0 on August 5, Seedance 2.5 on August 11 and Ideogram 4.0 with Grok Imagine 1.5 Video on August 13. The API changelog puts Seedance 2.5 and Grok Imagine Video 1.5 on August 7. Neither is wrong so much as differently scoped. But if you’re rebuilding a timeline for an audit, say which changelog you used. The unlimited Hailuo 3.0 promotion for new Max plans carried an August 23, 2026 expiry, and the pricing page no longer shows it.
seedance2_5exists, dated August 7 in the API- App and API changelogs date arrivals differently
- The Hailuo promotion expired on August 23
An HDR pipeline, not just a model
Two entries on the same day. A new model, ruby, converts SDR video into HDR, and Runway says it accepts uploads or generations from any model (paraphrase). The API pricing page bills it at 20 credits per second, and at 40 credits per second when the source is larger than 4 megapixels, which is roughly 4K. Separately, Gen-4.5 and Aleph 2.0 gained HDR output formats. Runway’s product changelog tags the Ruby entry Max and above, which is the detail a Pro subscriber needs before promising an HDR deliverable from the app.
ruby: 20 per second, 40 above 4 megapixels- HDR output on Gen-4.5 and Aleph 2.0
- Tagged Max and above on the product changelog
WAN 3.0 and Muse Image
WAN 3.0 arrived on Runway Dev as wan3. Runway describes it as generating up to 30 seconds of video with native audio, from a text prompt or a start-frame image. Pricing is 5 credits per second at 480p, 10 at 720p and 20 at 1080p. Muse Image arrived the same day as muse_image: text to image with optional reference images, prompts up to 4,000 characters and eight aspect ratios, at 1 credit per image. One trap. Muse Image is API-only, despite reading like an app-side model. It doesn’t appear on Academy’s models-and-pricing table at all, which is the app-versus-API split running in the direction buyers rarely check.
wan3: 30 seconds with native audio- Resolution is the cost lever, 5 to 20 credits per second
muse_imageis API-only, not app-side
ACEScg OpenEXR delivery
HDR outputs can now be delivered as scene-referred ACEScg OpenEXR sequences, and the API changelog entry names both ACES 1.3 and ACES 2.0 variants. If that sentence means nothing to your team, it isn’t for you. If it means something to your colorist, it’s the most consequential thing Runway shipped in August. Pin the identifier before you write a delivery spec: hdr_exr_acescg_sequence_1_3 or hdr_exr_acescg_sequence_2_0. Runway’s Professional outputs page describes the 2.0 one as the same scene-referred ACEScg delivery, inverted through the ACES 2.0 Output Transform.
- Scene-referred ACEScg
- ACES 1.3 and ACES 2.0 identifiers
- Aimed at finishing pipelines
MiniMax H3 Max
A new identifier, h3_max, with durations from 5 to 15 seconds, billed at 5 credits per second at 480p and 8 at 768p. Read it carefully. This isn’t hailuo3, which is still live, and neither rate matches the MiniMax H3 rates Academy shows on the app side.
h3_max, nothailuo3- 5 at 480p, 8 at 768p
- Academy quotes different rates again
GPT Image 2.5
Two identifiers, gpt_image_2_5_flare and gpt_image_2_5_sunburst, taking up to 16 reference images across 30 aspect ratios, with an output count of 1 to 10. Billing runs from 1 to 76 credits per image. At 1K and 2K: 1 for low, 5 medium, 16 high, 28 xhigh and 63 max. At 4K or auto: 2, 11, 19, 34 and 76. Add 1 credit per reference image, then multiply by the output count.
- Two quality families, one model
- Up to 16 references, 30 aspect ratios
- Cost scales with quality and output count
Ruby gains alpha and a plate-referenced build
On September 11 Ruby began preserving the alpha channel of a source video on the video-to-HDR route, writing alpha as the EXR A channel and using ProRes 4444 for ProRes output. On September 12 the two ACEScg EXR sequence outputs on /v1/video_to_hdr, hdr_exr_acescg_sequence_1_3 and hdr_exr_acescg_sequence_2_0, became plate-referenced, built from the source plate with colorimetry metadata updated to match. That changes the August 31 behavior described above. So if your colorist tested against the August behavior, retest.
- Alpha survives the HDR conversion
- ACEScg sequences built from the plate
- The August 31 description is superseded
A second finishing model
A brand-new model, enhance_frame_rate, retiming a video to a target frame rate, billed at 1 credit per 2 seconds. Ten targets: 24, 25, 30, 48, 50, 60, 120, 23_98, 29_97 and 59_94. Inputs can be at most 300 seconds. The drop-frame three are the point. If a deliverable has to land at 23.98 or 29.97 for broadcast, this is the model that gets it there. It’s the second post-production conversion tool Runway has shipped in a month, after Ruby, which reinforces the pattern: Runway is buying its way into the finishing stage, not only the generation stage. Watch the app side, though. The product changelog entry advertises a shorter list of targets than the API carries, and tags it Max+ Plans.
enhance_frame_rate, 1 credit per 2 seconds- Ten target frame rates, inputs up to 300 seconds
- Max+ Plans in the app
Five more API models and a draft mode
September 28: seedance2_5 gained a draft mode. A 480p preview bills at the 480p rate, and you can enhance it to 1080p through draftTaskId within 7 days, billed again at the 1080p rate. September 30: eleven_v4, for scripts up to 2,500 characters, at 2.2 credits per 1,000 characters until October 12, 2026 PT, then 5 per 1,000. October 1: grok_imagine_1_5_lite at 2, 3 and 14 credits per second for 480p, 720p and 1080p, and Runway says its 1080p is rendered at 720p and upscaled. October 6: nano_banana_2.1 at 4, 6 and 12 credits per 1K, 2K and 4K image, with up to 14 references. October 12: ideogram_4_5 at 1, 4, 8 and 25 per image by quality, and ideogram_4_0 at 4, 7 and 11.
- A draft you enhance is billed twice
- Budget
eleven_v4at 5 per 1,000 characters - Grok Imagine 1.5 Lite 1080p is an upscale
Kling on Runway Dev
Seven identifiers in one entry: kling2.5_turbo_pro, kling3.0_standard, kling3.0_pro, kling3.0_4k, klingO3_standard, klingO3_pro and klingO3_4k. Kling 3.0 runs 3 to 15 seconds. Kling O3 takes up to 7 reference images and video-to-video inputs up to 10 seconds. Multishot runs through providerSettings.multiPrompt, and cfgScale, from 0 to 1 with a default of 0.5, is exposed on 2.5 Turbo Pro and 3.0. Audio defaults to on, and Kling 3.0 Standard and Pro bill the no-audio rate only when audio is false. Write that flag into every call that doesn’t need sound.
- Seven Kling identifiers
- Multishot and up to 7 references on O3
- Set
audioto false to pay the lower rate
The web app
Generation, Edit Studio, Studio and Agent, with the roster Runway Academy prices. The broadest bench, including Kling 3.0 Turbo and Motion Control, which the REST API can’t reach, and the place most evaluation happens.
- Widest model roster
- Academy is the price list
- Where a wrong assumption starts
The REST API
A different list, not a shorter one. It carries identifiers that never appear in the app, like muse_image, wan3_prime and kling3.0_4k, and leaves out app-side models from LTX 2.5 to Topaz 4K. Since October 13, 2026 it reaches seven Kling models. It runs on its own money too: credits bought per project at $0.01, in a pool separate from any plan unless you’re on Enterprise, with five spend tiers that set concurrency and daily caps.
- Different roster, not a subset
- Seven Kling models since October 13, 2026
- Separate credit pool outside Enterprise
Runway MCP
The agent path inside chat tools. The page at runway.com/mcp lists Gen-4.5, Seedance 2.5, GPT Image 2, Gen-4 Image, Kling 3.0, Veo 3.1, Nano Banana Pro and Gen-4 Turbo. The product changelog added Workflows on August 20 and Ideogram 4.5 on October 2. HDR conversion on MCP is tagged Max+. It runs from ChatGPT, Claude.ai, Claude Desktop, Grok Bot, Cursor and Replit. No separate API key, and Runway says MCP uses your Runway credits, just like the app. The pricing page names it as an entitlement on Pro and Max.
- Kling 3.0 on its published list
- Billed to plan credits, no separate key
- Listed on Pro and Max
Runway Dev MCP
Launched September 2, 2026, and a different product from the consumer MCP. A hosted MCP server connecting a Runway Dev account to Claude, ChatGPT, Codex and Cursor, so an agent can research which models a project can use along with prices and capabilities, configure model routing and debug an integration.
- Developer account, not plan credits
- Claude, ChatGPT, Codex and Cursor
- Model research, routing and debugging
The Adobe plugin
Released September 8, 2026 for Premiere Pro and After Effects. Free to install on macOS and Windows, generation on a Standard plan or higher, billed to plan credits. Runway’s plugins page lists nine models: Gen-4.5, Seedance 2.5, Kling 3.0 Pro and Veo 3.1 for video; Gen-4 Image, Nano Banana and GPT Image 2 for stills; Aleph 2.0 and Runway Ruby for restyle and HDR. Runway’s help article calls it the same lineup as your plan on web.
- Nine models listed, Kling 3.0 Pro included
- Standard plan or higher
- Ruby’s plan gate in the panel isn’t stated
The DaVinci Resolve plugin
Released September 23, 2026 and the newest door. Needs DaVinci Resolve Studio 19 or later. Free installer, generation on paid plans, billed to plan credits. Aleph 2.0 restyles with up to five keyframes and results land in the Media Pool and on the timeline. Runway’s plugins page names Resolve as a host of the same panel and its nine-model list, but no page says every model on it works inside Resolve.
- Resolve Studio 19 or later
- Up to five keyframes on a restyle
- Shared plugin list; open the panel to confirm
From a clip generator to a campaign engine
The biggest change at Runway in 2026 isn’t a model. It’s Agent. Runway describes it as a chat-based marketing agent that builds and scales the work: concepts, ads, full campaigns and variants. Agent Skills, shipped in July 2026, collapse those jobs into single commands. So the human job moves. From prompting, to briefing and reviewing. That’s a different skill, right? And most teams haven’t trained anyone for it.
A marketing agent, not a prompt box
Agent works up the direction with you, produces finished video, and builds and scales campaigns across platforms, formats and markets from the same concept.
- Concepts, ads, campaigns and variants
- Multi-format and multi-market output
- Chat-based, so the brief is the interface
Agent Skills
Runway’s framing: build an ad campaign, create a commercial, localize your ads and more, with one command. Skills turn repeated campaign work into named, callable jobs instead of freehand prompting. Agent has moved since July, though only the changelog says so. Projects and Brand Kits came to Agent on August 11, 2026, tagged for Enterprise, and the new Team plan bundles up to 20 agent skills per workspace. The product page itself is still undated.
- Named jobs, repeatable output
- Projects and Brand Kits in Agent, August 11, 2026
- Up to 20 agent skills on the Team plan
Projects and Brand Kits reach Agent
Runway’s product changelog records Projects and Brand Kits arriving inside Agent on August 11, 2026, tagged Enterprise. It matters more than it sounds. An agent that can see the brand system and the project it’s working in is an agent whose output needs less correcting. It moves Agent from a standalone generator toward something that inherits your workspace context. The changelog is a rolling page rewritten in place, so confirm the entry and its plan tag before you plan around it.
- Agent inherits brand context
- Tagged Enterprise on the changelog
- Read the changelog entry yourself
Agent on the new Team plan
Team is the first self-serve tier to put real numbers on Agent capacity. Up to 20 agent skills. Shared Brand Kits across the workspace. Agent Connectors for Figma, Dropbox and Notion. Up to 100 shared projects, and a shared credit pool of 6,900 credits per seat per month. If Agent is why you’re buying Runway, this is the tier to price. It’s also the one a buyer is most likely to miss, because the pricing page keeps it behind a Team and Enterprise tab rather than in the default Individual table.
- Up to 20 agent skills
- Connectors for Figma, Dropbox and Notion
- Behind a tab on the pricing page
Brand Kits for everyone, and cheap drafts
Three more Agent entries on the product changelog. Agent Brand Kits on September 10, tagged All Plans, which puts brand context in Agent on every plan. Agent in Asset Comments on September 15, tagged Enterprise. And Seedance 2.5 Draft Mode in Agent on October 2, tagged All Plans, for fast 480p drafts you enhance after you’ve picked a winner. The draft habit is the one to build. Look at cheap drafts, then pay for the finals.
- Agent Brand Kits on all plans
- Agent in Asset Comments for Enterprise
- Seedance 2.5 drafts at 480p in Agent
Runway Ads
A new product built on Runway Agent that runs the creative side of paid advertising. It generates ad creative, publishes approved variants to Meta, Google and TikTok, reads performance back and regenerates. Runway says human approval is on by default. Keep it on. An ad loop that publishes and regenerates by itself is the campaign loop below with the review step deleted. The launch post gives no pricing, plan or availability details.
- Generates, publishes, measures, regenerates
- Human approval on by default
- No published price or plan
The controls that separate direction from luck
Most people treat a generator like a slot machine. You don’t have to. When brand, character or composition matters, these are the features that make a shot repeatable instead of a lucky roll.
References
Reference-driven characters and locations, so they stay consistent across shots and scenes. This is the key to brand and narrative continuity.
- Consistent characters
- Reusable locations
- Brand-safe recurrence
Layout Sketch
Draw on a blank canvas, or over an image, to control composition and place new elements before you generate.
- Compose deliberately
- Guide placement
- Less trial and error
Camera and motion brush
Say how the camera and individual objects move through a scene, rather than hoping the model guesses what you meant. Hope isn’t a shot plan, right?
- Camera paths
- Per-object motion
- Intentional movement
Keyframes
Set first, middle and last frames to control how a shot evolves across its duration. Kling O3 extends the idea with timed keyframe guidance and multi-shot output with audio, up to 4K. Since October 13, 2026 it’s on the REST API too, as klingO3_standard, klingO3_pro and klingO3_4k, with multishot through providerSettings.multiPrompt and up to 7 reference images. Kling 3.0 Motion Control didn’t make that move. It sits in the app. So decide which door you’re building on before you design a workflow around it.
- Start and end control
- Timed guidance in Kling O3, now on the REST API
- Motion Control stays off the API
Negative prompts
On the Veo 3.1 API models you can pass an optional negative prompt of up to 1,000 characters, naming what must not appear, instead of overloading the positive brief. Write everything against veo3.1 and veo3.1_fast. The older veo3 was sunset on August 4, 2026 and is gone.
veo3.1andveo3.1_fast- Text-to-video and image-to-video
- Added July 10, 2026
Aleph 2.0 editing
Change one frame and Aleph 2.0 changes the rest of the video to match, up to 30 seconds. This is edit-grade transformation, not a filter, and you see an image preview first.
- Prompt-based edits
- Preview in Edit Studio
- Up to 30 seconds
HDR and ACEScg finishing
Since August 20, 2026 Gen-4.5 and Aleph 2.0 output HDR directly. ruby converts existing SDR footage to HDR at 20 credits per second, and 40 above 4 megapixels. Since August 31, HDR output can be delivered as scene-referred ACEScg OpenEXR sequences, with both an ACES 1.3 and an ACES 2.0 identifier published. September refined it further: Ruby preserves a source alpha channel from September 11, and from September 12 the ACEScg sequences are built from the source plate. Three things to settle before the run rather than after. The delivery format, because a conversion pass is a second bill and Gen-4.5’s HDR formats carry a surcharge of their own. The ACES version, because there are two. And the plan, because Runway lists HDR, ProRes and image sequences as studio-grade features of Max, tags the ProRes and PNG sequence export entry Pro+ Plans, and tags Ruby Max+ Plans. In the app, Pro gets the export formats and not the converter. On Standard, none of this card is in reach.
- HDR output on the flagship models
rubyfor footage you already have- Ruby needs Max in the app, so check the plan first
Slow handheld close-up of a ceramic coffee cup on a wooden cafe table, morning light through large windows, shallow depth of field, soft filmic grade, gentle steam rising. Five-second product hero shot for a social ad, 9:16.
Use these character and location references for every shot in this sequence. Keep the character’s face, wardrobe and the set consistent across all six clips; vary only the action and camera.
Take this existing 8-second clip and change the time of day to golden-hour sunset while preserving the subject’s motion and the camera move. Don’t alter the composition.
Before the final run, tell me the plan: how many seconds at which model and resolution, whether audio is on, whether we are delivering SDR or HDR and what the finishing format or conversion adds, the estimated credit cost including any per-generation minimum, and which shots are cheap tests versus final deliverables.
For each shot in this list, recommend the single best model currently available in Runway and why, choosing among Gen-4.5, Veo 3.1, Seedance 2.5, WAN 3.0, MiniMax H3, Kling 3.0, Kling O3 or another, given the motion, realism and audio needs. Flag any recommendation that can’t be called from the REST API, such as Kling 3.0 Turbo or Kling 3.0 Motion Control, and name the door that can reach it.
Assemble the approved clips in Studio in this order, match durations to the storyboard, and export a 4K master plus a 9:16 social cut. Tell me whether the delivery spec needs HDR or an ACEScg EXR sequence, and which ACES version, before we render.
Before this ships, list for each asset: which model generated it and through which surface, the app, the REST API, either MCP server, the Adobe plugin or the Resolve plugin; whether we are attributing on that surface and in what form; whether any likeness, voice or brand asset needs written consent; and which plan tier the finishing format we promised actually requires.
What each plan buys, and what an API credit costs
Runway sells two kinds of credit, and on a self-serve account they never mix. App plans bundle a monthly allowance with features, and the features gate more than the credits do. API credits are sold per project in the developer portal at $0.01 each, with a $10 minimum payment to start. Runway’s developer docs say web app credits and API credits are separate pools. Enterprise is the exception: since September 18, 2026, Enterprise credit commitments are fungible across the web app and Runway Dev. So price a plan for its gates and the API for its rates. Plan figures come from runway.com/pricing, read October 15, 2026. Team figures come from Runway’s September 4, 2026 announcement.
| Plan | Annual, per month | Billed monthly | Includes |
|---|---|---|---|
| Free | $0 | $0 | 125 one-time credits, not monthly; 5GB storage; a selection of models. A trial seat |
| Standard | $12 | $15 | 625 credits per month; 20GB storage; 5 parallel generations; up to 3 projects; no watermarks; 4K upscaling. No Runway MCP, ProRes, image-sequence export or Ruby |
| Pro | $28 | $35 | 2,250 credits per month; 100GB storage; 15 parallel generations; up to 5 projects; 1 Brand Kit; 1 custom voice clone; top-up credits; Runway MCP; ProRes and PNG sequence export. No Ruby or Enhance Frame Rate in the app |
| Team | $55 per seat | $69 per seat | 2 to 9 seats, launched September 4, 2026. 6,900 credits per seat per month, pooled, with monthly rollover; 1TB storage; up to 100 shared projects; shared Brand Kits; up to 20 agent skills |
| Max | $76 | $95 | 9,500 credits per month, unused credits roll over 1 month; 500GB storage; 20 parallel generations; up to 10 projects; 3 Brand Kits; 3 voice clones; Runway MCP; HDR, ProRes and image sequences; Ruby; Enhance Frame Rate |
| Enterprise | Custom | Custom | Contract terms; no published figures. The only plan with the no-training commitment, IP indemnification and model licensing. Credits fungible across web app and Runway Dev since September 18, 2026 |
| Unlimited | Not sold | Not sold | Legacy subscriptions only; moves to Max on November 30, 2026. Annual refund option only before that date |
| API model | Credits per second | Watch for |
|---|---|---|
| Enhance Frame Rate (enhance_frame_rate) | 1 per 2 seconds | Added September 17, 2026; a retime pass billed on top of the generation, at any resolution |
| Grok Imagine 1.5 Lite (grok_imagine_1_5_lite) | 2 at 480p, 3 at 720p, 14 at 1080p | Added October 1, 2026; 1080p is rendered at 720p and upscaled |
| Gen-4 Turbo | 5 | The cheap seat for look-finding |
| MiniMax H3 Max (h3_max) | 5 at 480p, 8 at 768p | Added September 3, 2026; a distinct model from hailuo3, which is still live |
| WAN 3.0 (wan3) | 5 at 480p, 10 at 720p, 20 at 1080p | Up to 30 seconds with native audio; resolution is the whole cost lever |
| WAN 3.0 Prime (wan3_prime) | 6.8 at 480p, 14 at 720p, 28 at 1080p | Priced on the API with no Academy row |
| Kling 3.0 Standard (kling3.0_standard) | 9 without audio, 13 with | Added October 13, 2026; audio defaults on, so the lower rate needs audio set to false |
| Gen-4.5 | 12 | Flagship rate; 10 seconds is 120 credits, or $1.20. ProRes or PNG sequence adds 5; 10-bit or HDR adds 20, or 40 above 4 megapixels |
| Kling 2.5 Turbo Pro (kling2.5_turbo_pro) | 12 | A flat rate; Academy charges 15 at 1080p |
| Kling 3.0 Pro (kling3.0_pro) | 12 without audio, 17 with | Audio defaults on |
| Kling O3 Standard and Pro (klingO3_standard, klingO3_pro) | 13 and 17 | Up to 7 reference images; video-to-video inputs up to 10 seconds |
| Veo 3.1 Fast (veo3.1_fast) | 15 with audio, 10 without | The cheap half of the Veo pair, so price it before defaulting to full Veo 3.1 |
| Seedance 2 Mini | 16 | 64-credit minimum per generation |
| Ruby (SDR to HDR) | 20, or 40 above 4 megapixels | A two-step conditional, not a range; 40 for a 4K plate. Tagged Max+ in the app; no plan gate stated on Runway Dev |
| Seedance 2.5 (seedance2_5) | 20 at 480p, 30 at 720p, 68 at 1080p, plus half as much again on input | 80-credit minimum; Academy quotes the same tiers. A 480p draft enhanced to 1080p is billed at both rates |
| Aleph 2.0 (aleph2) | 28 | 56-credit minimum per generation |
| Seedance 2 Fast (seedance2_fast) | 29 | The cheaper Seedance 2 variant, often missed when a budget is built from the headline model |
| Seedance 2 (seedance2) | 36 at 480p and 720p, 40 at 1080p, 150 at 4K | The 4K rate is more than four times the 720p rate |
| Veo 3.1 | 40 with audio, 20 without | Audio doubles the rate |
| Kling 3.0 4K and O3 4K (kling3.0_4k, klingO3_4k) | 43 | No Academy row for kling3.0_4k; Academy shows 4K only under Kling O3 |
Habits that protect your credits and your brand
The biggest wins here are behavioral, not technical. Treat generation as a budgeted, staged process. Design for the edit, not for the one magic clip. And never assume a hosted model, a per-second rate or a plan price will still be there next quarter. Eight habits do most of the work.
Storyboard before you spend
Lock the shot list and the look before you commit credits to long, high-quality runs.
Test cheap, then commit
Find the look with Turbo, short durations or a 480p Seedance 2.5 draft. Upgrade to Gen-4.5 only for the shots that ship.
Think in shots, not films
Design 3 to 10 second clips you’ll cut together later. Don’t force a whole narrative into one run.
Pick the model per shot
Match Veo 3.1, Veo 3.1 Fast, Seedance 2.5, WAN 3.0, Kling 3.0 or O3, MiniMax H3 or H3 Max, or Gen-4.5 to the shot, instead of defaulting to whatever opened. Then check which door the winner is reachable through before you build on it. The app, the REST API, Runway MCP, Runway Dev MCP, the Adobe plugin and the DaVinci Resolve plugin don’t share one model list.
Budget = seconds × model
Duration times the model’s per-second credit rate is your cost lever, so plan runs against it. Three traps. Some models carry a per-generation minimum, including 80 credits on seedance2_5. Kling 3.0 on the API bills audio by default unless you turn it off. And finishing is a second bill: Ruby costs 40 credits per second, not 20, once the source is larger than 4 megapixels, and Gen-4.5’s HDR formats add 20 to 40 on top of its 12.
Exploit references
Use References and Layout Sketch whenever brand or character consistency matters.
Edit, don’t regenerate
When you almost have it, reach for Aleph 2.0 instead of rolling a fresh generation. A fix costs less than a reroll, right?
Diarize the dates, then re-read them
Runway published the August 4, 2026 Veo 3 sunset in advance and executed it on the day. The July 30 removal of gen4_aleph and gen3a_turbo was recorded on the day it happened, with no advance date. Kling reached the REST API on October 13 with no advance date either. And the Unlimited migration moved three months by an edit to a help article on August 27, 2026. So a date in your diary is a reminder to re-read the page, not a fact. The pages to watch are the API models page, the API pricing page, both changelogs and any help article carrying a date. The product changelog carries no history, so keep your own copy. And name a fallback for every hosted model in a standing workflow.
Range and speed don’t equal cleared, finished or affordable
Range is the product. Range is also the risk. Credits burn fast. Models get retired, sometimes with a date and sometimes without. Plans get retired on dates that themselves move. The API covers a different set than the app. Finishing formats are gated by tier. And generated people and claims carry rights and truth obligations. Some of that has an answer now. Read who the answer covers, though. Uncapped IP indemnification, no-training commitments that reach third-party providers and per-model admin controls are Enterprise terms. On a self-serve plan, Section 4.4 of the Terms of Use lets Runway train on your inputs and outputs. So sort each risk three ways: answered, answered only for contract customers, or not answered at all.
Your work trains the model unless you’re on a contract
This is the risk most Runway buyers never price. Section 4.4 of the Terms of Use, last updated September 15, 2026, says inputs and outputs may be used to train and improve Runway’s AI models. Labeling, classification and content moderation are named too. No opt-out setting is published on any first-party page. No paid tier carves itself out. Free, Standard, Pro, Max and Team are all inside it. The reverse commitment exists in exactly one document, the Enterprise Services Terms, last updated June 1, 2026. They say Runway may not use Customer Content as training data for the Services, and that the customer owns all rights, title and interest in it. So if a client contract, an NDA or a works council agreement forbids your material entering a training set, a credit card subscription doesn’t satisfy it. Get on a contract, or keep that material out of Runway.
· Source: Runway Terms of Use, Section 4.4, last updated September 15, 2026, and Runway Enterprise Services Terms, last updated June 1, 2026
A published date isn’t a promise, and the help center proves it
The Veo 3 sunset in the API docs held to the day. The Unlimited to Max migration in the help center didn’t. On August 27, 2026 Runway edited the article, deleted every mention of September 1 and put November 30 in its place. Same vendor, same quarter, two different levels of reliability. Technical deprecation notices in the developer docs have earned trust. Commercial dates in help articles haven’t. So weight a date by the page it sits on, right? A docs sunset is a plan. A help-center date is a draft.
Two Runway pages, two answers, and one of them prices a dead model
The Veo 3 version of this contradiction resolved. Two replaced it, and the new one is worse. Runway Academy, read October 15, 2026, still lists Gen-4 Aleph at 15 credits per second, and Runway’s own API changelog declared that model no longer available on July 30, 2026. That’s 77 days of a purchase price on a model another first-party page formally retired, counted to October 15, 2026. There’s a second instance inside the developer docs. Read October 15, 2026, the billing page at docs.dev.runwayml.com renders pricing for gen3a_turbo alone, switched off on the same July 30 date, and it has dropped out of the docs navigation while staying live at its URL. The canonical API price list is the pricing guide, and that one is correct. So don’t treat any single vendor page as the vendor’s position. And don’t assume a contradiction you resolved last month stays resolved.
Hosted models can be removed, and advance notice isn’t guaranteed
Sora 2 and Sora 2 Pro were removed on April 3, 2026. Gen-4 Aleph and Gen-3 Alpha Turbo went on July 30, 2026. The notice you get isn’t uniform. Runway gave Veo 3 an August 4, 2026 sunset date in its API model guide and executed it on that date. The July 30 pair is the other case: the only evidence for it is the API changelog entry of July 30 itself, recording that both models are no longer available. That’s a record of a removal already done, not a date published ahead of it. So there are two failure modes, not one. The first is a date published on a page nobody on the buyer side is watching. The second is a removal you learn about afterward. A standing watchlist of vendor deprecation pages works against the first, but only if somebody reads it. Against the second it shortens the gap rather than closing it. One evidence note. Runway publishes no dated notice of the Sora removal. As of October 15, 2026 it shows only in the absence of any Sora identifier from the API models page, the API pricing page and Academy. A hosted model is a vendor relationship, not a permanent feature, so name a fallback anyway.
Six surfaces, and the model list changes by door
Kling shows how fast this moves. It reached Runway MCP on May 27, 2026 and the REST API with seven identifiers on October 13. Kling 3.0 Turbo and Motion Control still have no API identifier. So the question was never app versus API. It’s which of six surfaces: the web app, the REST API, Runway MCP, Runway Dev MCP, the Adobe plugin and the DaVinci Resolve plugin. The editor plugins share one published list of nine models on runway.com/plugins, and Ruby is on it even though the app gates Ruby at Max. The API runs the other way too: muse_image, wan3_prime and kling3.0_4k have no Academy row. Sora is on none of them. Confirm coverage on the surface you’ll actually ship from.
Deprecations
Models move, and one of them has moved further than its own documentation suggests. The old gen4_aleph isn’t deprecated in favor of aleph2. It’s gone. Runway’s API changelog recorded gen4_aleph and gen3a_turbo as no longer available on July 30, 2026, and neither appears on either API page. Anything you standardize on can be retired. Rosters move the other way too, and just as quietly: Kling sat off the REST API for months, then arrived on October 13, 2026 with no date published ahead of it. No new deprecation or sunset was published on either API page or in the API changelog between August 19 and October 15, 2026, which is a quiet window rather than a policy change. The changelog does list aleph2_alpha as a deprecated alias, dated June 2, 2026. Keep a migration path, and read both changelogs: the API changelog for identifiers and dates, the product changelog for which plan tier a feature lands on.
Credit burn, now with a second bill
Long, high-quality, high-resolution runs eat credits fast. Some models also carry a per-generation minimum on top of the per-second rate, including 80 credits on seedance2_5 and 56 on aleph2. Kling 3.0 on the API bills its audio rate by default. Then finishing adds its own line items. ruby is billed at 20 credits per second, and 40 once the source is larger than 4 megapixels, on footage you’ve already paid to generate. Gen-4.5’s HDR formats add 20 credits per second, or 40 above 4 megapixels. And enhance_frame_rate adds another 1 credit per 2 seconds. One conversion that helps: if you bill through the developer portal rather than a plan, a credit costs $0.01, so 100 credits is a dollar and a 10-second Gen-4.5 shot is $1.20. Test cheap, lock the look, and decide the delivery format before you commit.
Plan and region gates
Some models and features are paid-only or limited by region, and since August the gating reaches the finishing pipeline. Runway lists HDR, ProRes and image sequences as studio-grade features of Max. Its product changelog tags ProRes and PNG sequence export Pro+ Plans, and tags Ruby, HDR conversion on Runway MCP and Enhance Frame Rate Max+ Plans. So in the app Pro can’t run the SDR to HDR pass or retime footage, and Standard can’t deliver an HDR master at all. Whether those gates follow Ruby onto Runway Dev, or into the editor plugins, isn’t stated anywhere. Verify availability before you promise a client a specific model or a specific format. Plans move too, and their dates move. Unlimited can’t be bought, legacy subscriptions hold until November 30, 2026, and that date replaced September 1 five days before September 1 arrived. The Team plan is a discovery risk of its own, because the pricing page keeps it behind a Team and Enterprise tab.
Likeness and voice
Act-Two and Characters capture performance and can clone voices. Get explicit consent and clear rights before anything ships.
The published answer to the rental problem
A hosted third-party model is a rental. The mitigation is stronger than most buyers assume. Runway’s Enterprise FAQ on third-party models states that third-party models run under the same enterprise terms, data protections and security standards that govern your account. It states that Runway has secured contractual commitments from all third-party model providers that they won’t train on customer content, inputs and outputs included. It states that the indemnification obligations in your enterprise terms cover third-party models on Runway. And it states that admins can enable and disable individual third-party models, with no data sent to a provider whose model is disabled. And on the question every procurement team asks about ByteDance and Kuaishou models, it states that Runway’s providers for Seedance and Kling models have confirmed that they will not process your customer data in China. That doesn’t stop a model being withdrawn. It does mean the legal and data exposure of using one is contracted rather than inherited. It’s an undated page, so have counsel read the actual terms.
· Source: Runway Help Center, Enterprise FAQ on third-party models, carrying only a Last Updated April 29th, 2026 stamp, read October 15, 2026
Ownership and commercial use
Runway’s usage rights help article states that, as between you and Runway, you retain ownership and all your rights to content you upload and generate on Runway. There’s no tier differentiation. Commercial use is permitted across plans, including monetized and non-monetized video uploads and product advertising. Runway’s enterprise page repeats full ownership of your outputs and adds that Runway doesn’t train on your data, but that page is selling the Enterprise contract. Don’t read it as a product-wide promise. Two limits to keep in mind. Ownership is not the same as no training: Section 4.4 of the Terms of Use lets Runway train on content you own, and owning an output doesn’t keep it out of the training set. Section 4.4 also grants Runway a non-exclusive, irrevocable, perpetual, worldwide, royalty-free, transferable and sublicensable license to inputs and outputs. And ownership as between you and Runway isn’t clearance against a third party. It settles the vendor question, not the rights question on a likeness, a track or a trademark.
· Source: Runway usage rights help article, which carries no date at all, read October 15, 2026, and Runway Terms of Use, Section 4.4, last updated September 15, 2026
Attribution: a style note that gets read as a rule
Runway says there’s no formal requirement to credit it in work made in the product. The developer attribution page, read October 15, 2026, is weaker than most summaries of it: when attributing to Runway, please use a Powered by Runway mark and link to Runwayml.com from the user interface. That’s how to attribute, not a stated duty to attribute, and no penalty is named. Runway supplies the lockups. Neither page is dated. So decide your own position rather than inheriting one, and if it carries commercial weight, put it in the contract.
An app-side estimate can still miss the API bill
Academy and the API pricing page agree on Seedance 2.5: 20, 30 and 68 credits per second at 480p, 720p and 1080p, plus 10, 15 and 34 per second of input video. The gaps sit elsewhere. Kling 2.5 Turbo Pro is a flat 12 per second on the API and 12 or 15 by resolution on Academy. The API’s kling3.0_4k has a rate of 43 and no Academy row at all. And Gen-4.5’s finishing formats carry a surcharge the headline rate hides: 5 credits per second for ProRes or a PNG sequence, 20 for 10-bit or HDR, and 40 above 4 megapixels. A 4K HDR Gen-4.5 second costs 52 credits, not 12. Build cost models from the surface you’ll bill on.
Model licensing is a different contract, not a bigger plan
Runway’s August 20, 2026 enterprise post announces a licensing program. Runway delivers closed model weights and a proprietary training and inference framework to a customer, who hosts the model in their own environment and fine-tunes it on their own data. Treat that as a separate commercial shape, not a premium tier. Different negotiation. Different security review. Different exit position. And no published price. If the reason you hesitate over Runway is that a hosted roster can change under you, this is the conversation to have, and it starts with sales rather than with a checkout page.
Which Runway page you read decides what you believe
Runway publishes across runway.com, runwayml.com, Academy, two changelogs, a developer docs site and a help center, and freshness varies between them. The product changelog at runway.com/changelog is active, with its newest entry dated October 2, 2026 when read October 15, and every entry tagged with the plan tier it applies to. But it carries no history, and its per-month entry counts don’t hold still, so a count from it is a snapshot and not a record. The problem isn’t that some pages are dead. It’s that the same fact is stated differently in different places. Two changelogs date the same model arrivals differently. Academy and the API pricing page disagree on Kling 2.5 Turbo Pro. Academy still prices a model the API retired. The docs billing page says credits are bought for an organization, and the pricing guide says for a project. And the domain question has an answer with a catch: runwayml.com returns a 302 Found to https://runway.com/. A 302 is temporary. So runway.com is canonical in practice while Runway hasn’t committed to it with a permanent redirect, and a saved runwayml.com link is one whose destination Runway can change without breaking anything.
| Task type | Minimum check | Human responsibility |
|---|---|---|
| Client deliverable video | Full render watched at target resolution and delivery format; artifacts and on-screen text checked | Final-cut approval |
| Character or avatar | Likeness rights and voice consent confirmed | Legal sign-off |
| Hosted third-party model | Offered on the plan and region in use, and reachable on the automation surface in use | The tool decision |
| Agent-built campaign | Every claim, statistic, logo and localized line fact-checked | Brand-owner approval |
| Credit-heavy run | Seconds × per-second rate, plus any per-generation minimum, audio rate, finishing surcharge and HDR conversion pass | The budget |
| Client, NDA or unreleased material | Enterprise contract confirmed before upload; every self-serve plan sits under the Section 4.4 training terms | The confidentiality call |
| Anything shipped through the API | An attribution decision on record; Runway publishes a lockup and a link format, with no stated requirement or penalty | The attribution decision |
| An HDR or ProRes deliverable | Plan tier and ACES version confirmed before the spec is agreed; in the app ProRes is Pro+ and Ruby is Max+ | The format promise |
| Anything with a vendor date on it | Date diarized, source page re-read on the same cadence, dated copy of the row kept | The migration window |
Where these facts come from
Runway ships weekly, retires models on real dates and moves commercial dates by editing support articles in place. Most entries below are dated first-party articles. Four are rolling pages, kept because each is Runway’s only first-party record of its subject: API credit pricing, API rate-limit tiers, the app-side rate card and the plugin model list.
aleph2 is the only Aleph identifier on the API pages↗Runway AgentRunway · undated product page · the chat-based marketing agent, Agent 2.0 and Agent Skills · records none of the dated Agent entries in the product changelog↗Kling O3Runway · undated product page · timed keyframes, multi-shot with audio, up to 4K · available on all paid and Enterprise plans · on the REST API as klingO3_standard, klingO3_pro and klingO3_4k since October 13, 2026↗Runway CharactersRunway · real-time avatars, launched March 2026 · undated product page · latency and frame-rate figures are vendor claims↗Introducing Runway GWM-1Runway Research · the general world model behind Characters, and the lineage behind the September 3, 2026 GWM Worlds 2 post↗Does the Max gate on Ruby apply to the API?
The product changelog tags Ruby and Enhance Frame Rate Max+ Plans, and the Ruby entry says Ruby is available on Runway Dev. The Professional outputs, API pricing, setup and tiers pages state no plan requirement. Open until Runway states it.
Is Ruby gated inside the plugin panel?
Runway’s plugins page lists Runway Ruby, and the plugin help articles require a Standard plan or higher for the plugin as a whole. Neither page says Ruby needs Max inside the panel, as it does in the app.
Does every plugin model work in Resolve?
Runway’s plugins page names DaVinci Resolve as a host of the shared nine-model panel. Runway publishes no Resolve-specific model list, and the Resolve help article says only that it carries the same lineup as your plan on web.
Did the Eleven v4 price rise take effect?
The API pricing page still reads 2.2 credits per 1,000 characters until October 12, 2026 PT, then 5 per 1,000. No Runway page reports the switch as done. Budget at 5.
Stale billing page or rendering fault?
The developer billing page renders pricing for gen3a_turbo alone, a model retired on July 30, 2026, and it no longer appears in the docs navigation. Nothing on the page separates stale content from a rendering fault. The pricing guide is the canonical list, and it’s correct.
Are changelog entries added late?
The product changelog carries no history, and its per-month entry counts don’t hold steady between readings. The page itself can’t tell you whether entries are backfilled or removed, so count only what you screenshot.
The Sora 2 Pro model page
No Sora identifier appears on the API models page, the API pricing page or Academy. Runway publishes no dated notice of the Sora removal. Whether the old Sora 2 Pro model page still answers with a 404 is unconfirmed as of October 15, 2026.
Which year do the banners mean?
The Enterprise “Sign by 10/31” banner and the Max “until October 18” Wan promotion print no year and link to separate terms. Read those terms before you count on either.
Team figures sit off the pricing page
Runway’s pricing page shows a Team & Enterprise tab with nothing under it. Team figures come from Runway’s September 4, 2026 announcement, the dated record.
No cross-reference between rosters
LTX 2.5, FLUX 3, Magnific Style Transfer, Fish Audio, Lyria 3, Topaz 4K, the legacy GPT Image models, and Kling 3.0 Turbo, Motion Control, 2.6 Pro and 2.5 Turbo carry Academy rates and no API identifier. Runway publishes no cross-reference, so the size of the app-only bench is unpublished.
AI Mindset