
Kandinsky 6.0 Video vs Runway Gen-4.5: One of Them Has Nothing to Say
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 90 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 96 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0717 tok/s
- OrcaNEWOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 88 tok/s
- OrcaNEWOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 133 tok/s
- OrcaNEWOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$4.00 / $20.00 per 1M tokens · 72 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 130 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 65 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 92 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 43 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 193 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 157 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
- OrcaOrca: OrcaCyber Zero 1.52026-10-10$3.00 / $7.50 per 1M tokens · 119 tok/s
- openaiOpenAI: GPT-6.1 Sol2026-09-2955Intelligence
- anthropicAnthropic: Claude Sonnet 5.52026-09-2855Intelligence
- typesafeTypeSafe: Jev 1.132026-09-24$0.04 / $0.00 per 1M tokens · 221 tok/s
Here is a comparison that can be settled by a single specification, and it is not a quality specification. Runway Gen-4.5 does not generate audio. Not as a limitation of one mode, not as an optional add-on — the model renders picture and nothing else, and every audio decision downstream of it is yours. Kandinsky 6.0 Video, open-sourced by Sber's Kandinsky Lab on 6 October 2026 under MIT in a Lite 3B and a Pro 29B size, generates 44 kHz audio with lip-sync in the same pass as the picture, and treats that joint generation as its entire reason for existing. So before any argument about motion quality or price, the matchup splits on whether the deliverable needs to arrive with sound attached. If it does, there is no comparison to make. If it does not, the rest of this page is the comparison.
Where the two do overlap
Set audio aside and the overlap is narrower than the differing reputations suggest. Neither model plans shots. Runway Gen-4.5 renders per clip with no multi-shot storyboard mode, and Kandinsky 6.0 Video has a single continuous take — five seconds, 121 frames at 24 fps, no cut concept, no extension mode in the release. Neither will hand you a sequence. Both are asked for one shot at a time and both are joined by you.
Both are also driven the same way: a text prompt or a still image, with iterative regeneration as the actual working method. Runway Gen-4.5 has been in market since December 2025 with an asynchronous API and a UI that a great many people have already built habits around. Kandinsky 6.0 Video arrived on 6 October 2026 with checkpoints and a diffusers integration, which means the loop is a Python script instead of a browser tab — different ergonomics, same shape of work.
What separates them past that is length, resolution, container and the meter. Clip length is the visible one: Runway Gen-4.5 covers 2 to 10 seconds per generation, so a longer beat is still one request where Kandinsky 6.0 Video would need two five-second renders and a join. That is a real advantage for establishing shots, holds and anything with a slow camera move, and it exists on the visual side of the ledger where Kandinsky's published evidence is weakest.
The credit meter, converted into a clock
Runway Gen-4.5 bills 12 credits per second of output, which at list is $0.12 a second — a ten-second 720p clip for about $1.20. Two extras sit on top of that and are worth knowing before you budget: ProRes or PNG export adds 5 credits a second, and HDR adds 20 more, which becomes 40 credits a second above roughly four megapixels. Those are the numbers you multiply when the deliverable is a grading-ready file rather than a share.
Kandinsky 6.0 Video has no rate at all, because there is nothing to buy. What it has is a hardware cost you supply, and the report is unusually forthcoming about it: a Pro Full HD five-second clip takes roughly 402 seconds of working time on an H100, 854 on an RTX 5090 and 1,247 on an RTX 4090, block offloading is required below a 72.8 GiB peak allocation, and the 16 GB VRAM preset quantises the text encoder to NF4 — the only preset change the report says alters the clip itself. Both models run at 24 fps, with Runway also offering 25, and both work across multiple aspect ratios.
Put the two side by side as clocks rather than currencies. Runway Gen-4.5 sells you roughly twenty seconds of wall-clock render time for a dollar. Kandinsky 6.0 Video sells you nothing and costs you six to twenty-one minutes of one GPU for every five seconds you keep — against which you set the amortised price of the card, the electricity, and the fact that you can render without a network connection or a rate limit.
• Length — Runway Gen-4.5: 2–10 s per generation. Kandinsky 6.0 Video: 5 s fixed at 121 frames.
• Audio — Runway Gen-4.5: none generated. Kandinsky 6.0 Video: 44 kHz with lip-sync, generated jointly.
• Resolution and container — Runway Gen-4.5: 720p at 24/25 fps, with ProRes or PNG export available. Kandinsky 6.0 Video: SD at 512×768, Full HD 1920×1080 via a built-in super-resolution stage, no intermediate format.
• Shot structure — Runway Gen-4.5: one clip per request, no planner. Kandinsky 6.0 Video: one continuous take, no planner.
• Weights — Runway Gen-4.5: closed. Kandinsky 6.0 Video: MIT, Lite and Pro, code and diffusers included.
• Evidence — Runway Gen-4.5: ten months of public output and third-party use. Kandinsky 6.0 Video: a vendor-run evaluation and a benchmark run against two comparators, neither reproduced externally.

The best-sourced objection to Kandinsky's own claim
Runway does not need to publish a comparison to have a strong position here, because the Kandinsky report already states its own limits. Its announced constraints, in its own words: clips of up to five seconds; base generation at SD resolution with Full HD obtained through super-resolution; and a gap remaining to the strongest proprietary systems in visual quality and overall audio quality. Runway Gen-4.5 is one of the systems that sentence is pointing at.
Set against that, the report's own results. In side-by-side human evaluation — pairs from the same prompt, anonymised clips, randomised left/right placement, shuffled task order, audio off then on, a tie option, majority-vote aggregation, roughly 200 or more pairwise comparisons per criterion — Kandinsky 6.0 Video Pro is competitive with Kling 2.6 and Veo 3.1 Fast with strengths that differ by criterion, and trails MiniMax H3 and Dreamina Seedance 2.0 on visual criteria while staying competitive on speech quality and audio-video synchronization. Reinforcement learning cut the word error rate of generated speech by 47%. On VABench it leads LTX 2.5 on speech quality and lip-sync accuracy and yields to it on VLM-judged alignment and expressiveness.
Runway appears nowhere in that list, in either direction. There is no head-to-head, no shared benchmark and no independent board that scores both. Anyone publishing a winner between Runway Gen-4.5 and Kandinsky 6.0 Video on quality is extrapolating from two unrelated evidence bases, one of which is ten months of public output and the other of which is a report by the model's authors.
The pipeline question, which is where routers earn their place
Neither of these models is routed on OrcaRouter, and that is a statement about our catalogue rather than about either product: Runway Gen-4.5 is served through Runway's own API, and Kandinsky 6.0 Video is a checkpoint you download and run. A pipeline built on either one, though, spends most of its calls on text rather than on pixels.
Consider what a five-second audio-video pipeline actually needs. A shot note has to become a long structured caption of the kind these models were trained on. Dialogue has to be written before a lip-sync pass can attempt it. Six candidate generations of the same beat have to be reviewed and one selected, and that review is a text call about a video. Then there is the part that decides whether the whole thing is worth running: the join decisions, the continuity notes, the script for the next clip. All of it is text, and all of it runs on one OpenAI-compatible endpoint across 200-plus models at provider list prices passed through at 0% markup, with a routing DSL that lets the cheap reviewer and the careful reviewer be different models on the same call site. Model fusion helps in a specific way here — a caption written by one model and critiqued by another tends to catch the invented details that a single pass waves through. And because a render loop is long, automatic failover is the difference between a failed caption costing you one call and costing you the clip.
The decision, stated as a question about sound
If the finished piece needs dialogue, ambience or lip-sync that arrives with the picture, Runway Gen-4.5 is not in the running and the only remaining choice is which audio-video model you deploy. That is a short list, and Kandinsky 6.0 Video is the open entry on it.
If sound is handled elsewhere — by a separate model, a voice actor, a library — then Runway Gen-4.5's ten-second clips, mature API and exporting options are the stronger working tool today, and its evidence base is the older and thicker of the two. The thing to watch in that case is not a benchmark. It is whether Runway closes the audio gap, because that single specification is currently the entire structural argument for putting a brand-new open checkpoint into a production pipeline ahead of a model with ten months of use behind it.


Where the caption and review calls live, model fusion helps catch invented detail that a single pass waves through.
