Sora 2 Leaves the API on 24 September, and the Replacement Column Is Empty — What That Does to Your Cost Per Video

Thirteen days before the shutdown, a developer on Hacker News wrote the sentence this whole situation deserves: "After being burned by the rug-pull of OpenAI retiring the Assistants API in favor of Responses last year, I swore off using heavily stateful APIs for language model access." (Hacker News, z2, 2026-09-11.)

The thing he is describing has a table now. Pulled first-party on 2026-09-21 from developers.openai.com/api/docs/deprecations, the entry reads:

"On March 24th, 2026, we notified developers using the Videos API and Sora 2 video generation model aliases and snapshots of their deprecation and removal from the API on September 24, 2026."

Underneath it, a table with three columns — Shutdown date, Model / system, Recommended replacement — and six rows: the Videos API itself, sora-2, sora-2-pro, sora-2-2025-10-06, sora-2-2025-12-08, sora-2-pro-2025-10-06. Every one of them dated 2026-09-24. And the Recommended replacement cell for all six reads three dashes.

The empty column is data, not a formatting slip

It would be easy to dismiss a blank cell as a page that nobody finished. The same page rules that out. Of the 35 deprecation entries on it, the ones that have successors say so: the most recent entry, dated 2026-09-11, retires gpt-5.4-cyber on 1 October and names gpt-5.6-cyber in the replacement column. Another row points an older model at gpt-5.6-terra.

So the column works. It gets filled when there is something to put in it. Six consecutive blanks next to demonstrably filled rows elsewhere on the same page is a statement, and the statement is: for this capability, there is no in-house successor to point you at. We are not going to speculate about why — motive is not in the table, and everything written about it this week that sounds confident is someone guessing. What is in the table is enough.

Everyone is answering the wrong question

Search this and the first page is almost entirely blog posts and tool sites comparing alternative video models — we counted eight of them plus a Wikipedia entry, and not one page from OpenAI itself. They all arrive at a shortlist. That is a fine article to write and a useless one to read, because the model you pick is the cheap part of this decision. The expensive part is what a finished, publishable clip costs you afterwards — and that number is not the number on anyone's pricing page.

Four reasons the headline rate misleads:

1. The billing units are not the same shape. Per second of output, per generation, per credit, per GPU-minute, per seat. Two of those scale with clip length and three do not. You cannot rank them until you convert every one into cost per finished clip at your duration and resolution.
2. Takes per usable clip is the hidden multiplier. If one model gives you something shippable on the second attempt and another needs five, a rate that looks 40% cheaper is 50% more expensive. Nobody publishes this number because it depends on your prompts and your standards — so you have to measure it yourself, on your own briefs, before you commit.
3. Audio, captions and aspect ratio may or may not be included. A model that returns silent 16:9 needs a voice track and a re-frame before it is a vertical short. Those steps have their own meters.
4. The swap itself costs engineering hours, and they are charged to the week you can least afford them — the week the old endpoint stops answering.

Runway, writing first-party on 2026-09-10 about where video generation is heading, put the underlying economics in one sentence: "We consistently find that cost per output at a given quality bar determines viable use cases." And the mechanism behind it: "when models run faster, they take up less GPU time, making them more cost effective." (runway.com, Towards Instant Video Generation, checked 2026-09-21.) Which is the honest framing — cost per output at a given quality bar, not cost per second, and not the number on the card.

You are not comparing models. You are comparing the total price of one clip you would actually post — and every vendor has chosen a unit that makes that hard to compute.

The real lesson is about the slot, not the occupant

Strip the brand names out and this is a dependency story, and the people who have lived through it said it best. From a thread in July:

"That's not really true though, providers are deprecating models and I have at least 10 emails to prove it." — Hacker News, TurdF3rguson, 2026-07-11

And the reply, which is the whole exit strategy in one line: "A provider deprecating a model doesn't mean the .gguf file disappears from my computer." (Hacker News, trollbridge, 2026-07-11.) A third commenter had already put a clock on it two weeks later: "anything you built 3 months ago is probably going to break soon. AI solutions need better insurance around model deprecation." (Hacker News, albert_e, 2026-07-24.)

So here is the design principle, stated without reference to any vendor. If the video step in your pipeline has exactly one slot with exactly one occupant, then a line in somebody else's changelog is a production incident. Insurance, in this context, does not mean picking the biggest provider. It means the swap being a configuration change rather than a rewrite.

That is a structural argument, and it happens to be how our own video side is built — several generation routes behind one workflow, including ones that do not call a hosted generator at all, so a model going away changes which route runs rather than whether anything runs. It is also why we keep writing about metering units rather than prices: see what you are actually buying from a video generation API, and the wall nobody prices, and three categories of repurposing tool with three completely different meters.

A checklist for the next three days, and for the next time

1. Find every call site. Not just the generation call — the polling, the retry handler, and anything that parses a response shape specific to that API.
2. Write down your quality bar before you shop. Resolution, duration, whether audio is required, acceptable takes per usable clip. Without this you will compare rates instead of costs.
3. Price two candidates in your own unit, running the same three briefs through both, and record attempts as well as spend. This is the step that produces a real number.
4. Keep the second candidate wired. Not running — wired. The cost of a spare slot is one configuration entry; the cost of not having one is the week you just spent.
5. Subscribe to the deprecations page, not the launch blog. Launches are marketed to you. Shutdowns are published where you have to go looking.
6. Check whether the platform you publish to already gives you a generator before renting one — we mapped that for short video in where the free options actually stop.

FAQ

What exactly stops working on 24 September 2026?

According to the deprecations page checked on 21 September 2026, six entries share that shutdown date: the Videos API itself, and the model aliases and snapshots sora-2, sora-2-pro, sora-2-2025-10-06, sora-2-2025-12-08 and sora-2-pro-2025-10-06. The announcement text states developers using them were notified on 24 March 2026. The consumer-facing Sora application was discontinued earlier and separately; this is the API-side removal, and they are two different events.

What is the recommended replacement?

The table does not name one. All six rows show three dashes in the Recommended replacement column. On the same page, deprecations that do have successors list them by name — a September entry retiring gpt-5.4-cyber points at gpt-5.6-cyber. So the blank is meaningful rather than missing. Anyone telling you the official replacement is a specific model is not reading the table.

How should I compare video models on cost?

Convert everything into cost per finished clip at your own duration, resolution and quality bar, then multiply by the number of attempts it actually takes you to get one you would post. Vendors meter in incompatible units — per second, per generation, per credit, per GPU-minute — so headline rates are not rankable. Add any separate charges for audio, captioning or reformatting, because a silent horizontal clip is not a vertical short yet.

If you only do one thing this week: open the pipeline and count how many places name a specific model. If the answer is more than one, that is not an architecture, it is a set of hostages — and the ransom note is published, on a public page, months in advance, in a column you now know how to read.