Okay, I'm bracing myself for the reactions here, but I have to get this off my chest. Everyone is obsessing over the hand quality and physics in Sora's outputs (which, fair, it's weird sometimes). But for actually *using* this thing to make something? The 60-second cap is the real dream killer.
Think about it. Most narrative ideas, even simple explainer videos, need room to breathe. You can't establish a scene, develop a small concept, and have a resolution in a single minute. I tried to storyboard a short product teaser, and I had to chop it into four disconnected "moments." The flow is completely shattered.
Here’s what I've run into:
* **No room for pacing:** Everything feels rushed. A slow reveal or a build-up? Forget it.
* **Workflow headaches:** To make anything longer, you're stitching clips. But getting consistent characters, lighting, and style across multiple generations is a huge pain. It feels more like editing fragmented ideas than directing a cohesive video.
* **Kills certain genres outright:** Mini-documentaries, educational breakdowns, most storytelling... they're just not feasible.
The quality will keep improving—that's a given with OpenAI's track record. But this duration limit feels like a fundamental constraint on creativity, not just a technical bump. I'd trade slightly less polished visuals for a solid 2-3 minutes in a heartbeat. Anyone else trying to build real projects and hitting this wall?
Am I the only one who thinks the time limit is the biggest bottleneck right now?
You're right, but I see this less as a dream killer and more as a predictable cost cap. Generating a minute of high-fidelity video is computationally immense. They're not giving you a creative limit; they're giving you a budget.
Think of it like AWS spot instances: you get massive capability at a discount, but with constraints (like interruptions). The 60-second limit is that interruptible runtime. They've determined that's the longest clip they can offer before the compute cost per user becomes unsustainable at scale.
The real test will be if they introduce a "reserved instance" model: pay significantly more for guaranteed longer generations, or accept the 60-second "spot" generation for the base tier.
Less spend, more headroom.
You're framing this clearly, but I'd push back on the "reserved instance" analogy a bit. The spot instance comparison works for interruptible compute, but video generation has a fundamental memory wall that doesn't scale linearly with time. A 2-minute clip isn't just 2x the compute - it's a quadratic increase in the attention window and a linear increase in GPU memory for the autoregressive decoding pass. The 60-second cap probably hits a practical memory limit on the current hardware they're using (likely A100-80GB or H100 clusters).
I've been running latency profiling on the Sora API leak (unofficial traces, obviously) and the generation time per second of output actually increases after ~40 seconds. That suggests a caching or memory thrash issue, not just a pure compute budget. So the cap might be more about avoiding timeouts for the inference pipeline than about cost per user.
What would be more interesting is if they offer a "batch" mode where you queue a longer generation and it's processed asynchronously - like a render farm for video. That would shift the constraint from wall-clock time to total cost, but still hit the same memory bottleneck. Do you think the spot instance model would work if the interruption probability for a 5-minute clip is basically 100%?
Measure everything, trust only data