
FastFrames Video Generation vs Google Veo 3.1: Real Time Against Finished Film
- 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
- 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 · 120 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
- 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
These two systems are not competing for the same job, and the fastest way to see that is to ask what each one refuses to do. FastFrames Video Generation, which Apple released on 16 November 2026 with the paper following on the 17th, will not give you a file you can hand to a client: it tops out at 480p, it is silent, and it has no hosted endpoint at all. Google Veo 3.1 will not give you a frame in under a second: the standard tier is priced per minute of finished output, and its latency is measured in the tens of seconds. One is infrastructure for something interactive. The other is a production house. Comparing them on quality alone would be comparing a camera to a film crew.
Six dimensions, and who measured each one
Start with what is measured, and be careful about who measured it. The Veo 3.1 figures below come from Artificial Analysis's video boards as captured in October 2026, before FastFrames existed. The FastFrames figures come from Table 1 of Apple's own paper, on the same prompts and seeds at 480p. No independent body has run FastFrames, and no board has scored it. Where you see a Veo number it is a third-party measurement; where you see a FastFrames number it is a vendor claim.
• Latency — FastFrames-S at 36 ms per frame on one B200, roughly 28 requests per second vs Veo 3.1 at tens of seconds per clip, sold by duration rather than by frame.
• Resolution — FastFrames at 832×480, no other option vs Veo 3.1 at 720p, 1080p and native 4K depending on tier.
• Audio — FastFrames generates none at all vs Veo 3.1 native 48 kHz stereo with dialogue, music and effects in the same call, with generate_audio defaulting to off.
• Price — FastFrames has no price because it has no endpoint; you buy GPUs vs Veo 3.1 at $24.00 per minute of output, or $9.00 on the Fast tier and $4.80 on Lite.
• Access — FastFrames as downloadable weights under Apple's AMLR licence, self-hosted on Linux with CUDA 12.8 vs Veo 3.1 through Google's API and third-party platforms, no weights.
• Provenance — FastFrames publishes none vs Veo 3.1 embedding SynthID watermarks and C2PA content credentials in every output.

What Veo 3.1 is actually buying you
Veo 3.1's pitch is not "best looking". It is "finished". The model takes a text prompt, up to three reference images, a seed and a negative prompt, and returns a 4, 6 or 8 second clip with synchronised 48 kHz stereo — dialogue, ambience and effects generated together with the picture, lip movement matched to speech. Scene extension chains those clips. Tiers add resolution up to native 4K. Every file carries a SynthID watermark and a C2PA credential, which is the difference between an asset a legal team will clear and one it will not.
On quality it is good and not leading. In the October 2026 capture of Artificial Analysis's image-to-video board, Veo 3.1 sat twelfth at Elo 1,082 with a ±7 interval across 7,040 samples. The Fast tier was seventeenth at 1,066 ±6 over 14,688 samples and the Lite tier fifteenth at 1,071 ±8 over 5,122 — three tiers, statistically indistinguishable quality, prices five times apart. On the text-to-video board it ranked nineteenth at 962 ±10 over roughly 3,000 samples. Read those together and the practical conclusion is that Veo 3.1's tier choice is a budget decision, not a quality decision, and that its real moat is the audio, the 4K and the provenance rather than the Elo.
What FastFrames is actually proving
The interesting claim in Apple's release is narrower than "faster video model" and much harder to dismiss. Autoregressive video diffusion is slow for reasons that have little to do with the transformer: the variational autoencoder that turns latents back into pixels dominates the budget, and so does the KV cache update that carries context into the next block. Apple replaced the first with a 63.87 million-parameter pixel decoder — three causal ConvNeXt blocks plus a linear head — and the second with a shared-stem, four-block refinement network costing 2 milliseconds on the small model and 6 on the large one. The DiT is frozen in the final stage.
• VAE decode — about 147 ms per frame becomes roughly 1 ms.
• DiT, 1.3B — 41.6 ms becomes 16.4 ms, or 5.8 ms spread across eight GPUs.
• DiT, 14B — 89.2 ms becomes 75.6 ms, or 20.3 ms on eight GPUs.
• Causal Forcing++ total, 1.3B — 271.8 ms becomes 35.8 ms, a 7.6× reduction.

Both checkpoints are distilled from Wan2.1 — FastFrames-S from Wan2.1-T2V-1.3B, FastFrames-L from Wan2.1-T2V-14B — using the Causal Forcing++ recipe, and the pixel decoder was trained for about a week on 32 B200s over 150,000 rollouts, with an adversarial pass on eight B200s against a frozen DINOv3 ViT-S/16 discriminator. The authors' VBench numbers put FastFrames-S at 84.23 and FastFrames-L at 84.44, against 84.11 for the baseline it was distilled from and 84.42 for Krea-Realtime, which runs at 1,675 ms per frame.
Two things FastFrames cannot be used for, and one it can
It cannot replace a delivered clip
480p is a hard ceiling in the released configuration. Apple reports 720p latency — 127 ms per frame on the small model, 116 ms on the large one across eight GPUs — but no 720p weights or settings ship. There is no audio path at all, so any talking-head, product demo with voiceover, or anything a client will watch with sound needs a separate audio pipeline. And at 12.8 seconds per clip for the 14B on a single GPU, the large model is slower than playback unless you own a node.
It cannot be called
There is no Apple endpoint. FastFrames is a research release: weights on Hugging Face under the AMLR licence, code on GitHub, and a LingBot-World v2 demo app under CC BY-NC-SA 4.0 — non-commercial. To use it you need Linux x86-64, an NVIDIA GPU, CUDA 12.8, uv sync, and a checkpoint directory. That is a procurement problem before it is a technical one.

It can make a live loop feel live
At 36 ms per frame, FastFrames-S on one B200 produces video 1.7 times faster than it plays back. That is the threshold where a generative preview responds to input within a frame or two, which is what makes streaming avatars, controllable world models and interactive prototyping different products rather than slower ones. Apple's own demonstrations take two existing systems interactive: LingBot-World's 1.3B variant goes from 1,085 to 52 ms per frame, a 16× drop, and its 14B from 3,862 to 233. RealWonder 1.3B goes from 1,141 to 51. None of those are Veo replacements either — they are things Veo cannot do at any price, because no hosted API bills by the frame in a session.
If you need to ship something this quarter
The decision rule is short. If the deliverable is a file — a commercial, a product video, an explainer with voice — Veo 3.1 is the finished tool and FastFrames cannot enter the conversation, because it cannot give you sound or 1080p. If the deliverable is an experience — a live demo, an interactive character, a tool where a human steers and sees the result immediately — no hosted video model does this today, FastFrames is the most credible open path to it, and the cost is a GPU and a licence review.
There is a third case, and it is the common one: a team that wants to find out whether the interactive case is worth building. Prototyping that against a hosted model is cheaper than buying B200 time, and OrcaRouter is where the hosted side is one key — the same endpoint covers Kling 3.0, Kling 3 Turbo and Kling Video O1 alongside MiniMax H3, all at provider list pricing with 0% markup, so a vendor rate change shows up on your invoice the day it happens. FastFrames itself is not in that catalogue and we do not serve it; nothing in this release changes what we host. What a single key buys is the ability to measure what latency your users actually tolerate before you commit hardware to beating it.
Where this goes next
The gap between the two systems is a gap in what silicon is doing during generation, and it will close from both ends. Hosted vendors are already selling faster tiers, and the pixel-decoder technique is model-agnostic in principle — it is trained per base model, but nothing about it requires Wan2.1. The version of this comparison that matters is eighteen months out: a hosted 1080p model with audio that returns a clip in under a second, priced per second rather than per minute. Until that exists, the honest summary is that FastFrames won the latency race in a category Veo is not running in.
