The Sora 2 API retires on 2026-09-24, which is 18 days away, and OpenAI lists no replacement. So the practical question for anyone shipping video generation is not which model is best but which supplier you want to be exposed to next time this happens. The three routes are direct provider APIs (cheapest per call, one relationship per model), aggregators like fal and Replicate (one integration, many models, a markup), and platform APIs that bundle a workflow. Disclosure up front: 8frame does not sell a raw generation API. We expose an MCP server, which is a different shape, and this guide says where each route wins including where we are the wrong answer.
TL;DR
- Sora 2,
sora-2-proand the Videos API stop working 2026-09-24. Its endpoints are already gone from at least one aggregator's catalog - Cheapest per call: direct provider APIs, at the cost of a separate integration, key, and billing relationship per model
- Fewest integrations: aggregators (fal, Replicate, and now Runway, whose API carries wan3, seedance, veo3.1, hailuo, gemini_omni_flash and more)
- Rough per-second rates today: Veo 3.1 around $0.40 standard and $0.05 lite, Gemini Omni Flash around $0.10, Wan 3.0 from $0.035, Runway Gen-4.5 at $0.12
- If you want a workflow rather than an endpoint: an MCP server or a canvas, which is what we sell and what a raw-API team should not buy
The three routes
Direct provider APIs
You integrate Google, ByteDance, Alibaba, xAI or Runway individually. Cheapest per generation because nobody is marking up, and you get features the moment the provider ships them.
The cost is operational: a key, a billing relationship, a rate-limit regime and a payload schema per provider, and a migration project every time one deprecates something. Which is exactly the Sora situation, and the reason teams reconsider this route in weeks like this one.
Aggregators
One integration, many models, and switching a model is a string change rather than a project. fal and Replicate are the established pair; Runway has quietly become a third, since its developer pricing now lists wan3, seedance, grok, hailuo3, veo3.1, gemini_omni_flash, seedream and more alongside its own models.
You pay a markup, you inherit their availability, and you are exposed to their catalog decisions. Note that aggregators delist ahead of provider shutdowns: Sora's endpoints disappeared from fal's catalog before the September date, so a model can stop being reachable through your aggregator while the provider still runs it.
Platform APIs and MCP
Higher-level: not a generation endpoint but a workflow, with storage, chaining, and often a UI attached. Our own MCP server sits here, letting an agent generate images and video without you building the pipeline.
Where this is the wrong choice: if what you need is a raw endpoint inside your own product with your own queueing and storage, buy an aggregator or go direct. A workflow platform is overhead you will fight.
What the rates look like
Per second of generated video, from provider documentation. Verify before you budget; these move.
| Model | Approx per second |
|---|---|
| Wan 3.0 (480p / 720p / 1080p) | $0.035 / $0.07 / $0.14 |
| Veo 3.1 Lite | $0.05 (720p) |
| Gemini Omni Flash | ~$0.10 at 720p |
| Runway Gen-4 Turbo | $0.05 |
| Runway Gen-4.5 | $0.12 |
| Veo 3.1 Fast | $0.15 |
| Veo 3.1 Standard | $0.40 |
Two things to note when budgeting. Audio is frequently priced separately and can double a silent rate. And several models bill by tokens rather than seconds, so the published per-second figure is a derived estimate rather than the billing unit.
The lock-in questions to ask before integrating
What is the deprecation policy? OpenAI gave six months. Not everyone will.
Are model versions pinnable? If your product's output changes silently when a provider updates a model, you have a support problem you cannot debug.
What happens to in-flight jobs at shutdown? Rarely documented, and it decides whether a migration is a sprint or an incident.
Can you dual-route? The cheapest insurance is a second provider integrated but not primary. Teams that had one after the Sora announcement spent an afternoon; teams that did not spent a month.
Migration checklist, if you are on Sora now
Map every call site, including anything parsing the Videos API response shape or assuming Sora's 4, 8, 12 second durations. Test parity on two vendors, not one. Re-check audio assumptions, since Sora generated it and several replacements price it separately. Then budget: the closest replacements land between $0.05 and $0.40 per second, so per-second cost is broadly comparable and your total will not be.
Full per-use-case migration table in Sora 2 API shuts down September 24.
FAQ
What is the best AI video API for a startup? An aggregator, almost always. The markup buys you the ability to change models without a migration, which at startup stage is worth more than the margin.
Does 8frame have a video generation API? Not a raw one. We expose an MCP server so agents and tools like Claude can generate on the canvas, which suits workflow integrations and does not suit teams building their own pipeline.
How do I avoid another Sora? Integrate two providers and keep the second warm. Everything else is optimism.
Pick the supplier relationship before the model, because the model will change. See how the field compares on identical prompts in best AI video generator 2026.