
FastFrames Video Generation vs Grok Imagine Video: Both Autoregressive, One of Them Ships
- 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 · 88 tok/s
- 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 · 90 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 · 119 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 · 143 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 · 919 tok/s
The closest thing to a technical sibling that FastFrames Video Generation has in production today is Grok Imagine Video, and the resemblance is not superficial. Both generate video autoregressively — each block conditioned on the frames that came before — and both use that structure to extend a clip rather than to render a fixed window. Apple's release, published on 16 November 2026 with the paper on the 17th, uses it to hit 36 milliseconds per frame on a single GPU. xAI's uses it to let you keep a shot running to fifteen seconds with sound. The difference is that one of these is a downloadable research artefact and the other has been on a public API since January 2026, and that difference — not the benchmark table — is what decides the matchup for anyone with a deadline.
What each one costs, in the units you would actually budget
FastFrames has no price, because there is nothing to buy. The weights are on Hugging Face under Apple's AMLR licence, and the only cost is hardware: one B200 for the 1.3B model or eight for the 14B to stay interactive, plus CUDA 12.8 and Linux. Grok Imagine Video is sold per second and reported by Artificial Analysis as a per-minute rate, which makes the arithmetic unusually clean.
• Grok Imagine Video, original release — $4.20 per minute of output, which is $0.07 per second. A 15-second clip is about $1.05.
• Grok Imagine Video 1.5 — $8.40 per minute, $0.14 per second, roughly $2.10 for the same fifteen seconds.
• Grok Imagine Video 1.5 Lite on the text-to-video board — also $8.40 per minute, which is a pricing oddity worth flagging rather than explaining away.
• FastFrames-S — no rate. One B200 at roughly $2 to $4 per GPU-hour, generating 28 clips of five seconds per second of wall clock in the theoretical limit, so the marginal cost approaches electricity.
• FastFrames-L — same, on eight GPUs, which is where the "cheap" argument gets harder.
The cost story only flips for FastFrames at volume, and only if you already own the GPUs. For a team renting, the crossover is far enough out that latency has to be the reason you choose it, not price.

Where Grok Imagine Video is simply ahead
Three shipped capabilities that FastFrames does not have in any form.
• Resolution — Grok Imagine Video outputs 480p, 720p or 1080p depending on the call; FastFrames is fixed at 832×480.
• Audio — xAI generates picture and sound together, with lip-synced speech and up to three preset voice references; FastFrames produces silent video and has no audio path.
• Control surface — up to 14 reference images, up to four mid-video keyframes plus first and last frames, and extensions chained from the last frame out to fifteen seconds; FastFrames takes a text prompt, and that is the entire interface.
On measured quality the Grok family sits mid-board rather than at the top, in the October 2026 Artificial Analysis captures: the original release was thirteenth on the image-to-video board at Elo 1,072 ±7 over 14,371 samples, Grok Imagine Video 1.5 ninth at 1,098 ±8 over 5,267, and the 1.5 Lite tier sixteenth on text-to-video at 993 ±10 over 5,715. So the 1.5 update bought about 26 Elo — a real improvement at the edge of the confidence intervals, not a step change — for double the price per second. Those are third-party numbers from October; FastFrames has no board entry to compare against, and Apple's own VBench totals of 84.23 and 84.44 for the small and large models are not commensurable with an Elo.
Where FastFrames is ahead, and it is only one axis
Speed, overwhelmingly, and the purpose that speed serves. Grok Imagine Video is a request-response system: you send a prompt, you wait, you get a clip. FastFrames-S on one B200 returns frames at 36 milliseconds each — 28 per second in aggregate throughput — which is faster than a 16 fps stream plays back. The 14B model needs eight B200s to reach 45 ms per frame; on a single GPU it drops to 158 ms and stops being interactive.
That is not a number that makes Grok look slow so much as a number that describes a different product. xAI is selling finished short clips with sound. Apple published a component for a live loop — and demonstrated it by taking other people's systems interactive, dropping LingBot-World's 1.3B variant from 1,085 to 52 ms per frame and its 14B from 3,862 to 233, with RealWonder 1.3B going from 1,141 to 51. None of those systems ship audio at 1080p either. The realistic read is that FastFrames is a technique waiting for a product, and the technique is good: replacing the VAE decode with a 63.87 million-parameter pixel decoder takes that stage from about 147 ms to about 1, and a small KV refinement network replaces the cache update for 2 to 6 ms.
The licensing trap nobody mentions in the benchmark table
FastFrames is open in a narrow, careful sense. The model weights carry Apple's AMLR terms. The LingBot-World v2 code and weights that ship alongside as a demo are CC BY-NC-SA 4.0 — non-commercial, share-alike. The remainder of the repository is under Apple's Sample Code Licence. If you were planning to drop the demo app into a commercial product, the licence stops you before the 480p ceiling does. Grok Imagine Video has no such ambiguity: it is an API, you pay per second, and the terms are a service agreement rather than a grant of rights in the weights.
That trade is worth naming plainly, because "open weights" gets used as shorthand for "free to use" and this release is not that.

How to pick between them
Ask what happens after the clip exists. If a human watches it once and moves on — a social post, a storyboard, an ad variant — Grok Imagine Video at $0.07 to $0.14 per second with audio and 1080p is the cheaper and better answer, and FastFrames cannot compete on output quality because it does not produce a finished output. If the clip is an input to something else — a preview that updates as a user types, a character that responds within a frame, a world model being explored — then per-clip pricing and a request queue are the wrong shape entirely, and FastFrames-S on one GPU is the only thing in this comparison that fits.
For the middle case, where you want to test whether the interactive shape is worth building before buying hardware, the cheapest route is to prototype against a hosted model you can call today. OrcaRouter covers the hosted video models on one key at provider list pricing with 0% markup — Kling 3.0, Kling 3 Turbo, Kling Video O1 and MiniMax H3 among them — and the flat pass-through means a vendor price change reaches your bill the same day instead of at renewal. Grok Imagine Video and FastFrames are both absent from that catalogue: we do not serve either, and nothing here implies otherwise. What one key does give you is a stable way to measure what latency your users will tolerate, which is the number that tells you whether Apple's release is relevant to your roadmap at all.
The honest summary
Grok Imagine Video is the better video generator by every measure except the one Apple optimised, and Apple optimised that one further than anyone else has. Two years from now the interesting question will not be which of these two is faster, because the hosted side will have caught up enough that the distinction stops mattering for most products. It will be whether anything hosted lets you steer a generation while it is still rendering. That is the capability FastFrames demonstrates, and no API in this comparison sells it.

