Versely Product
Versely Queue Generation: World's First Batched AI Video at 90% Less Cost (2026)
Sora 2 charges $1 per 10-second clip. Veo 3.1 Standard charges $7.50. Versely Queue Generation ships comparable footage at roughly $0.10 — and it is the first product to make batched LTX 2.3 a real workflow, not a research demo. Here is how it works.
If you wanted 100 ten-second AI clips this week, your credit card would care a lot about which model you picked. On Sora 2 base that bill is $100. On Veo 3.1 Standard it is $750. On Sora 2 Pro at 1080p it lands somewhere between $300 and $500 depending on length. On Versely's new Queue Generation lane, the same 100 clips cost roughly $10.
That is not a typo and it is not a coupon code. It is the difference between paying for on-demand inference and paying for batched inference on an open-weight 22B model that finally hit production quality this spring. As of June 2026, Versely Queue Generation is the only consumer-facing creator product that exposes batched LTX 2.3 as a first-class workflow — and the cost delta is large enough that it changes what kind of campaigns are economically possible. This piece is the deep dive: the economics, the architecture, the quality comparison, and where Queue Generation belongs in a real production workflow alongside premium models like Sora 2 and Veo 3.1.
The cost shock: 100 clips, five models
Before we get into how Queue Generation works, look at what it replaces. Here are the published per-clip costs for a standard unit of work — one ten-second 1080p clip with synchronized audio — and what 100 of those clips cost end of month.
| Model | Per-second | 10-sec clip | 100 clips | Audio in same pass |
|---|---|---|---|---|
| Veo 3.1 Standard | $0.75 | $7.50 | $750 | Yes |
| Sora 2 Pro (1080p) | $0.30–$0.50 | $3.00–$5.00 | $300–$500 | Yes |
| Runway Gen-4.5 | $0.15 | $1.50 | $150 | No (separate) |
| Veo 3.1 Fast (1080p) | $0.12 | $1.20 | $120 | Yes |
| Sora 2 base / Kling 3.0 | $0.10 | $1.00 | $100 | Yes / No |
| Versely Queue Generation (LTX 2.3) | ~$0.01 | ~$0.10 | ~$10 | Yes |
Sora 2 base sits at $0.10/sec, Sora 2 Pro 1080p at $0.30–$0.50/sec, and the consumer Sora app sunset in April with the API retiring September 24, 2026 (Sora pricing tracker, May 2026). Veo 3.1 Standard runs at $0.75/sec for 4K with native audio, and even Veo 3.1 Fast 1080p sits at $0.12/sec (AI video API pricing, April 2026). Kling 3.0 and Runway Gen-4.5 round out the $0.10–$0.15/sec tier.
Versely Queue Generation lands almost two orders of magnitude below the most expensive option and roughly 10x cheaper than the cheapest premium API — at quality that holds up against every model in that table for most non-hero use cases. The way it gets there is the part that matters.
Why batched generation is structurally cheaper
There is a very specific reason on-demand inference is expensive, and it is not greed. It is GPU economics.
When you submit a single prompt to Sora 2 or Veo 3.1, the inference provider spins up a slice of an H100 (or equivalent) to handle your request. That GPU has to load the model weights into VRAM, hold them there, and run your prompt. Then it either holds the weights warm for the next user, idling expensive memory, or it tears them down and re-pays the load cost when you come back. Either way, you are paying for memory bandwidth that is not generating frames.
The industry data on this is brutal. GPU utilization in generative inference workloads typically sits between 20–40% because so much of the chip's memory bandwidth is spent loading parameters rather than computing outputs (Anyscale on continuous batching, 2024). You are renting a Ferrari and driving it in a school zone.
Batched inference flips the equation. Instead of one prompt per GPU pass, the scheduler waits until it has N prompts ready, loads the weights once, and runs all N through the model in parallel (or in a tightly packed serial loop on the same warm pod). Real-world figures from continuous batching deployments hit 60–85% GPU utilization versus 20–40% on naive serving — that is roughly a 2–4x throughput multiplier on the exact same hardware at the exact same hourly rental rate.
For video generation, the multiplier is even bigger than for LLMs because video models are larger, weight-loading is more expensive, and the prompts are I/O-bound in ways that benefit from packing. When Versely's queue scheduler holds a job for 5–15 minutes to fill a batch window and then runs 32 prompts through a single warm LTX 2.3 pod, the per-clip compute cost collapses from the on-demand rate of roughly $0.04 down to closer to $0.008–$0.012. Add safety, storage and orchestration overhead and the user-facing price lands at ~$0.10 per ten-second clip — a number that wasn't physically possible to offer 12 months ago on this quality tier.
The trade is latency. Queue Generation jobs return in 5–30 minutes instead of 60–120 seconds. For hero shots that is a non-starter. For the 90% of creator video work that is variation testing, B-roll packs, social refresh, and ad creative batches, the latency cost is irrelevant and the savings are decisive.
Inside Versely Queue Generation: the architecture
Queue Generation is built on three layers that work together. None of them are novel in isolation. The combination, exposed as a public product, is what makes this the world's first batched AI video pipeline for creators.
Layer 1: LTX Video 2.3 as the engine
Lightricks released LTX Video 2.3 on March 5, 2026 — a 22-billion-parameter diffusion transformer with native 4K at up to 50 FPS, synchronized audio in a single forward pass, native vertical 1080×1920 composition, and an Apache 2.0 license for commercial use (LTX 2.3 launch overview). It is the first open-weight model that we benchmarked at hero-tier quality for everything except the most demanding cinematic and lipsync shots — and it is the model we built Queue Generation around.
For deep specs and the open-source ROI math, see our LTX Video 2.3 vs commercial models breakdown. The TL;DR for this piece is that LTX 2.3 is fast enough, sharp enough and licensed permissively enough that a batched lane is worth building around it.
Layer 2: The batch scheduler
The scheduler is the layer that makes the economics work. Every Queue Generation submission enters a shared pool. The scheduler watches three signals:
- Pool depth — how many jobs are waiting
- Window age — how long the oldest job has been waiting
- Pod warmth — whether an LTX 2.3 worker is loaded and idle
When pool depth crosses a threshold (typically 16–32 jobs) or window age crosses a cap (15 minutes), the scheduler dispatches the batch to a warm pod. Batches stay tightly packed: identical resolution and duration get bundled, audio-on jobs are grouped together, and identical model-version requests run back to back so weights stay loaded.
If you are a single user submitting one job, you ride along on someone else's batch window. If you are a content team submitting 80 jobs at once, you fill your own window and dispatch immediately. Either way, the per-clip cost is the same because the GPU utilization is the same.
Layer 3: The status pipeline
Batched inference creates a UX problem: how do you tell the user what is happening when their job is sitting in a 12-minute queue? Versely's status pipeline solves it with five stages that map cleanly to what the scheduler is doing:
- Queued — job is in the pool, not yet dispatched
- Scheduled — batch window closed, dispatch in progress
- Rendering — GPU is generating your specific clip
- Encoding — audio sync, final encode, watermark removal
- Ready — signed R2 URL returned, asset appears in your gallery
The whole flow is non-blocking. You can submit a job, close the tab, and a notification fires when the asset is ready. Multi-job batches show per-clip progress so you can watch a 50-clip campaign render across 8–12 minutes instead of staring at a single spinner. The same status object is what powers public template runs, story-to-video multi-scene jobs, and text-to-video submissions when you flip them into the queue lane.
Quality comparison: Queue Generation vs Sora 2 / Veo 3.1
The cost story is meaningless if the output is unusable. So we ran a head-to-head across six prompts that represent typical creator workloads.
| Prompt type | Sora 2 Pro | Veo 3.1 Standard | Queue Generation (LTX 2.3) |
|---|---|---|---|
| Product B-roll (handheld, kitchen) | 9.2 / 10 | 9.4 / 10 | 8.8 / 10 |
| Lifestyle stock (coffee shop, ambient) | 9.0 / 10 | 9.1 / 10 | 8.9 / 10 |
| Abstract motion graphic | 8.7 / 10 | 8.9 / 10 | 9.0 / 10 |
| Hero talking-head with lipsync | 9.5 / 10 | 9.6 / 10 | 7.4 / 10 |
| Complex camera move (orbit, parallax) | 9.3 / 10 | 9.5 / 10 | 8.2 / 10 |
| Vertical TikTok-native hook | 8.9 / 10 | 9.0 / 10 | 9.1 / 10 |
Internal ratings by three reviewers, 10-point scale, May 2026. Scores within 0.5 are inside reviewer noise.
The pattern is consistent with what creator teams report informally. Queue Generation matches or beats premium models on B-roll, stock, abstract, and vertical-native compositions — exactly the work most creators are generating most of the time. It loses meaningfully on hero lipsync and the most demanding camera moves, which is where Sora 2 Pro and Veo 3.1 still earn their per-second rates.
For a wider model comparison, our Sora 2 vs Veo 3.1 vs Kling 3 head-to-head covers the premium tier in more detail. The point of Queue Generation is not that it replaces those models. The point is that it removes the cost-per-clip floor for the 80–90% of work that does not need them.
Pricing breakdown: per clip, per minute, per campaign
Queue Generation is sold as credits, but the math is easier in concrete units. As of the June 2026 rollout:
| Unit | Sora 2 base | Veo 3.1 Fast | Queue Generation |
|---|---|---|---|
| 1 × 10-sec clip | $1.00 | $1.20 | $0.10 |
| 1 minute of video (6 clips) | $6.00 | $7.20 | $0.60 |
| 50-clip variant test | $50.00 | $60.00 | $5.00 |
| 100-clip ad campaign | $100.00 | $120.00 | $10.00 |
| 500-clip weekly refresh | $500.00 | $600.00 | $50.00 |
| 3,000-clip monthly refresh | $3,000 | $3,600 | $300 |
The 3,000-clip monthly refresh is the line that changes strategy. A content team running parallel batch creative ops at 100 assets per day is looking at $90,000+ per year on Sora 2 base for premium-tier output. The same volume on Queue Generation lands at $3,600 per year. The freed budget either expands campaign count, accelerates refresh cadence to fight the 9-day Meta creative-fatigue window, or simply drops to the bottom line.
For a wider economics teardown including agency comparisons, see our content cost vs agency breakdown.
When to use Queue Generation vs premium models
Queue Generation is not a replacement for Sora 2 or Veo 3.1. It is a different slot in the production workflow. The decision tree:
Use Queue Generation when:
- You need volume — 20+ clips per session, hundreds per week
- You are running variant tests (A/B/C/D versions of the same concept)
- The clip is B-roll, stock, motion graphic, or generic lifestyle
- You are generating vertical-native content for TikTok, Reels, Shorts
- Latency of 5–30 minutes is acceptable (it is, for almost everything)
- You are batch-generating UGC backgrounds, product cutaways, or transition footage
Use premium models (Sora 2 Pro, Veo 3.1 Standard) when:
- It is a hero shot anchoring a campaign
- You need premium lipsync on a talking-head
- You need a complex camera move (orbit, dolly, parallax) that has to read clean
- The asset will run at the top of a paid funnel where every frame is scrutinized
- You need it in 60 seconds, not 15 minutes
The pattern that wins is hero-and-volume: budget the premium model for the 5–10% of clips that are doing the heavy lifting, and route the other 90% through Queue Generation. The teams running this pattern in early access were producing 5–8x more total output for 30–40% of their previous monthly spend.
How to submit a queue job, step by step
The flow is intentionally close to the on-demand experience so existing Versely users don't have to learn a new tool. From the AI video generator or text-to-video page:
Step 1 — Compose the prompt. Same prompt box as on-demand generation. Add reference images if you have them. Pick aspect ratio (16:9, 9:16, 1:1) and duration (4s, 8s, 10s, 12s, 16s, 20s).
Step 2 — Flip the Queue toggle. Above the submit button there is a lane selector: Instant (premium model, 60-second latency, full per-second rate) and Queue (LTX 2.3 batched lane, 5–30 minute latency, ~$0.10 per ten-second clip). Flip to Queue.
Step 3 — Optional: stack the batch. If you have a variation matrix, hit "Add variant" to queue 5, 10, or 50 prompts at once. Each variant becomes a row in the status panel. The scheduler will pack them into the same batch window and you will pay the batched rate on every one.
Step 4 — Submit and close. The status moves to Queued. You will see an estimated batch close time (typically 2–10 minutes out). Close the tab. A notification fires when your assets are ready.
Step 5 — Review in the gallery. Ready clips appear in your asset gallery with the standard download / repurpose / auto-caption / video-to-shorts actions. You can chain the output into a slideshow or a story-to-video build without re-uploading.
That is the entire flow. For deeper workflow patterns built on top of Queue Generation, the batch generation for content teams playbook covers the parallel-batch operating system that takes full advantage of the new economics.
Why this is a defensible "world first"
The "world's first" claim deserves a justification because the category is crowded with hyperbole. Three things are true about Queue Generation as of June 2026 that are not true of any other consumer creator product:
- It is the first product to expose batched LTX 2.3 as a public lane. Other tools that route to LTX 2.3 do so synchronously, one prompt per pod, which means they pay the on-demand rate and pass it through. Versely is the only product running a public batch scheduler against shared LTX 2.3 pods.
- It is the only batched lane at this quality tier that is priced for individual creators, not enterprise contracts. Cloud inference providers offer batch tiers (OpenAI's Sora batch tier discounts 50%, for example), but those are reserved for high-volume API customers and still cost 5–10x what Queue Generation costs.
- It composes natively with the rest of the Versely toolchain. Queue jobs flow into the same gallery, the same public workflow templates, the same story-to-video builds, and the same agentic chat interface. The batched lane is not a separate app — it is a lane within the existing surface.
The defensibility comes from the orchestration layer. Building a batch scheduler is non-trivial. Building one that integrates with the rest of a creator workflow, handles failure modes, manages cold-pod problems, and surfaces a clean status pipeline is a year of engineering work. The cost advantage compounds for whichever product builds it first and we are happy to be that product.
FAQ
Q: How long does a Queue Generation job actually take?
The published SLA is 5–30 minutes per clip, but the typical wall time is 7–12 minutes for a single clip during normal-traffic hours. Batches of 20+ clips submitted together usually complete inside 15 minutes total — because the scheduler dispatches the full batch to a warm pod immediately rather than waiting for a window to fill. If you submit during peak traffic (weekday evenings, US Pacific), latency can stretch toward 25 minutes. We do not throttle or queue-skip paying users; the batched lane is the same lane for everyone.
Q: What happens if a clip fails in the middle of a batch?
Failed clips auto-retry inside the same batch window. If the retry also fails (rare — usually a content-policy block or a malformed prompt), the credit is refunded to your account and the failure surfaces in the status panel with a structured reason. You don't pay for failed generations and you don't lose your batch position.
Q: Can I mix Queue Generation and Instant in the same campaign?
Yes, and this is the recommended pattern. Generate your hero shots through Instant with Sora 2 Pro or Veo 3.1, then queue the supporting B-roll, cutaways, and variants through Queue Generation. Both flows land in the same gallery and can be assembled in the same edit, slideshow, or movie build. The hero-and-volume pattern described above is exactly this.
Q: Is the output watermarked or limited in any way?
No watermark on paid plans. You get the same 1080p output, same audio quality, same commercial-use rights as Instant lane generations. Free-tier accounts get a small Versely watermark; any paid tier (Creator, Team, Business) ships clean. License-wise, LTX 2.3 is Apache 2.0 and Versely's commercial-use terms pass that through.
Q: How does Queue Generation handle audio?
LTX 2.3 generates synchronized audio in the same forward pass — that includes ambient sound, music beds, and basic lip movement for talking-head shots. Audio is included at the same $0.10/clip price; there is no upcharge for the audio pass. For premium lipsync (sales VSL, brand spokesperson) you will still want to use lipsync video on top of an Instant-lane base, but for B-roll, lifestyle, and ambient work, the in-model audio is good enough to ship as-is.
The interesting thing about AI infrastructure is that the cost curve looks vertical when you are on it and flat when you look back at it. Sora 2 base at $0.10/sec felt cheap in March. Veo 3.1 Fast at $0.12/sec felt revolutionary at GA. Queue Generation at $0.01/sec is the next step on the same curve — and the one that finally makes high-volume AI video production look like high-volume AI text production.
If you are running a campaign right now, the math is simple. Use the premium models for the shots that have to land. Queue everything else. The freed budget is the difference between testing 5 hooks per week and testing 50 — and the team that tests 50 wins the cycle.
Open Queue Generation today inside text-to-video, AI video generator, or story-to-video. The first batch you queue tonight will probably be cheaper than the coffee you ordered to write the brief.