
FastFrames Video Generation vs Runway Gen-4.5: When 720p Meets 480p
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 89 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-079 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 · 134 tok/s
- OrcaNEWOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 87 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 121 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 68 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 91 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 44 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 · 195 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 118 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 187 tok/s
Set the two resolution figures next to each other and something useful happens to this comparison: it deflates. Runway Gen-4.5 generates at 720p, in two-to-ten-second clips, at 24 or 25 frames per second, at 12 credits per generated second — about $0.12 a second on the developer tier where a credit is a cent, with a documented five-credit-per-second surcharge for ProRes or PNG sequence export. FastFrames Video Generation generates at 832×480, 81 frames at 16 fps, in 36 milliseconds a frame on a single B200, silent, from a text prompt only. One of those ceilings is 720p and the other is 480p, and the gap between them is smaller than the gap between either and a 4K master. What is actually far apart is everything that happens before and after the render: Gen-4.5 is the generation half of a studio platform with its own editing, export and colour tooling; FastFrames is a pair of checkpoints and a licence file.
The two ceilings, honestly compared
Resolution is the axis where readers expect a rout and instead get a narrower question.
• Gen-4.5 — 720p generation, with a plan-level 4K upscale entitlement that is a separate operation and does not make the original generation native 4K. Download dimensions, not the export label, are what you compare.
• FastFrames — 832×480 fixed. There is no resolution setting, no 720p toggle and no upscale path in the repository. Apple reports 720p latency figures in the paper, but the released configuration is 480p.
• Frame rate — Gen-4.5 exposes 24 or 25 fps as a parameter, which is a genuine delivery requirement for broadcast and for anyone matching existing footage. FastFrames is fixed at 16 fps, and that is a property of the distilled model rather than a setting someone forgot to surface.
• Duration — Gen-4.5 gives two to ten seconds in one generation. FastFrames gives 81 frames, which is 5.06 seconds, and reaches longer runtimes by autoregressive chaining.
So: a 1.5× pixel-count difference, a frame-rate difference that matters more than the resolution one, and a duration ceiling that is roughly similar. If your deliverable is 720p social video at 24 fps, FastFrames is not a candidate — not because 480p is unusable but because 16 fps is, and because there is no path from 480p/16 to 720p/24 that does not involve inventing frames. If your deliverable is a live preview or a control loop where the frame rate is a property of the interaction rather than of the file, Gen-4.5's 24 fps parameter is irrelevant and its 720p ceiling is a cost you are paying for nothing.
What the credits are buying, and what the GPU bill is buying
Both models can be costed before you press generate, which puts them ahead of anything metered by tokens or by the model's own decisions. The units still do not line up.
• Gen-4.5 — 12 credits per generated second, flat, across a rate card where the output setting barely varies. A five-second generation is 60 credits, $0.60 at $0.01 per credit. Ten seconds is 120 credits, $1.20. ProRes or PNG sequence output makes it 85 credits, $0.85, for the five-second case. Every one of those figures is knowable in advance.
• FastFrames — no rate at all. The cost is one B200 for 2.9 seconds (small model, 5-second clip) or eight B200s for 3.6 seconds (large model). If the hardware is yours, that is a rounding error. If it is not, it is a reserved-instance contract priced in weeks.
• The take you discard — Gen-4.5 charges per generated second whether or not the take was usable. FastFrames charges the same GPU-seconds for a bad take as a good one, which is the cheaper failure mode by a wide margin and the reason exploratory workflows gravitate to per-GPU-second billing.
• The thing neither rate card tells you — accepted footage is not generated footage. Runway's own documentation distinguishes attempted generations from accepted output, and every budget built on the headline rate assumes a hit rate. Budget per delivered clip, including the takes you throw away.
Runway's pricing has a property worth naming because it is genuinely good: the flat per-second rate makes a 720p budget a one-line calculation regardless of the settings you choose. FastFrames has no price, so it has no such property — it has a hardware dependency instead, and hardware dependencies are budgeted annually rather than per project.

The platform is the part that is not on the spec sheet
Gen-4.5 does not arrive alone, and pretending it does would misrepresent what a buyer gets.
Video-to-video work goes through Runway's separate Modify Video tooling rather than through the model's own conditioning path, which means the editing workflow exists but is a different surface with its own behaviour. HDR profiles are documented on the developer side with their own surcharges. Native audio exists on the model, and separate Runway audio tools sit beside it. ProRes and PNG sequence export are documented and priced rather than hidden.
FastFrames has none of that, and it also has something Runway does not: the whole training pipeline. Apple published every training-stage checkpoint for the large model — the causal teacher, the consistency-distilled stage, the DMD-distilled stage, and the final generator with its pixel decoder and KV refinement network — plus a working inference path, a latency benchmark script that takes a GPU count, a VBench evaluation workflow and a demo app under apps/lingbot. That is a reproducible artefact, not a service. If your interest is in how a causal video stack can be made interactive, that artefact is worth more than any hosted endpoint; if your interest is in shipping a commercial on Friday, it is worth nothing at all.
The licence is where this stops being abstract. FastFrames' weights carry Apple's AMLR terms, the LingBot-World v2 demo components are CC BY-NC-SA 4.0 — non-commercial — and the rest of the repository falls under Apple's Sample Code Licence. Runway sells a commercial service with commercial terms. Anyone planning to build a product on FastFrames needs to read three licences before writing a line of code, and the non-commercial clause on the demo components is the kind of thing that gets discovered late and expensively.

What each side has measured, and who did the measuring
Runway describes Gen-4.5 on its own research page as its most advanced video generation model, with improvements in motion quality, prompt adherence and temporal consistency, available on all paid plans. That is a capability claim with no numbers attached. Gen-4.5 does not appear in the top twenty-eight rows of the Artificial Analysis image-to-video board — not a statement that the model is weak, but a statement about what Runway has chosen to submit, and it does change what kind of decision this is. There is no independent Elo for Gen-4.5 and no sample count.
FastFrames publishes numbers and they are all its own. VBench totals of 84.23 for the 1.3B and 84.44 for the 14B, against 84.11 for the Causal Forcing++ baseline it was distilled from at 272 milliseconds a frame. Unreproduced by anyone independent, on a benchmark that rewards the temporal smoothness autoregressive models are designed to produce, across a table whose entire spread is about two and a half points while the latency spread is four orders of magnitude. Vendor-reported, in the strict sense, and labelled as such everywhere it appears here.
So the two models have the same evidentiary status — unverified vendor claims — and that symmetry is itself the finding. A page that ranked them would be inventing a hierarchy out of two unevaluated datasets.
Which one, by job
Take Runway Gen-4.5 when the deliverable is a short, 720p clip that has to land inside a Runway workflow, when ProRes or PNG sequence export is a requirement rather than a nicety, when 24 or 25 fps is a specification you cannot move, or when the flat credit rate makes a project budget possible to write down. Accept the 720p generation ceiling and the upscale that is not the same thing as native 4K.
Take FastFrames when the interaction is the product and the file is a by-product — a live preview, a streaming avatar, a controllable world model, or a research programme that wants the technique rather than the output. Accept 480p, 16 fps, silence, no image conditioning and a licence review. The one thing you get that Gen-4.5 cannot offer at any credit rate is a frame every 36 milliseconds on a single card, and the honest framing of that is not "faster video generation" but "a different class of application."
OrcaRouter does not carry Runway Gen-4.5 or FastFrames — there is no route for either on the catalogue, and this piece is not claiming one. What one key does reach is the hosted video layer that already works today: MiniMax H3 at $0.08 per second at 768p and $0.13 at 2K, and the Kling 3.0 family, sitting next to 200-plus language models on an OpenAI-compatible endpoint at provider list price with 0% markup, so a vendor rate change is live here the same day and a degraded endpoint fails over rather than stalling a queue. For a team deciding between a credit-metered platform and a GPU-metered research release, that is the cheapest available way to get a real cost and latency baseline first.
What would resolve it
Runway submitting Gen-4.5 to the image-to-video board would replace the central asymmetry of this page with an actual number, and a generation tier above 720p would remove the one place these two models are genuinely close. On the other side, FastFrames needs a hosted endpoint before any preference board can measure it, and Apple has published weights without one.
Until either happens, the decision does not turn on quality at all. It turns on frame rate, on whether you need a file or a stream, and on whether your legal team would rather review a terms-of-service page or three licences at once.

