The Sora API will be shut down on September 24, 2026. Since the app and web versions ended in April, this is truly the end.
My name is Yuto, a company employee in my late 20s. This time, I’m talking about costs, not my impressions. I haven’t tried everything myself, so I won’t be writing about image quality. The prices are only those I was able to confirm on the official pages as of September 17, 2026.
First, the facts: What is stopping, and when?
I confirmed this on the OpenAI deprecation information page. Notification to developers was on March 24, 2026, and removal from the API is on September 24, 2026. This applies to the Videos API, sora-2, sora-2-pro, and their snapshots. The app and web versions (sora.com / iOS / Android) were already terminated on April 26, 2026.
What makes this different from a standard model replacement is that there is no mention of a recommended successor model. It just ends there.
Price per second (as displayed on official pages on September 17, 2026)
Starting with pay-as-you-go.
-
Veo 3.1 Lite (720p): $0.05 per second. 1080p is $0.08
-
Veo 3.1 Fast (720p): $0.10. 1080p $0.12
-
Veo 3.1 Standard (720p/1080p): $0.40. 4K $0.60
-
Seedance 2.5 (480p): approx. $0.2205. 720p is approx. $0.4730
-
Seedance 2.5 (480p with video reference): approx. $0.1323. Including video reference is 0.6x
The source for Veo 3.1 is the Gemini API pricing page, and for Seedance 2.5, it is the fal.ai model page. The “approx.” is because actual costs are calculated by tokens, and the displayed prices are estimates based on 16:9.
How much does one 10-second clip cost?
For Veo 3.1 Lite (720p), it’s $0.50; for Fast, it’s $1.00; and for Standard, it’s $4.00. Seedance 2.5 with 480p reference is about $1.32, and 720p is about $4.73. There is about a 9x difference between the cheapest and the most expensive. Looking only at this, Lite seems like the only choice, but the reason that isn’t the case is next.
The discarded portions are also part of the cost
You cannot use every clip you create. Unless you count the ones you discard, you won’t get the true cost.
So, let’s make one assumption: a 40% adoption rate, meaning 4 out of 10 generated clips are usable. This is not my measured result, but a hypothetical value set for calculation. Please replace it with your own numbers.
Under this assumption, a 60-second short would require 15 cuts = 150 seconds of generation, and a 90-second short would require 23 cuts = 230 seconds. With Veo 3.1 Lite (720p), one 60-second video is $7.50, one 90-second video is $11.50, and one per week (4 per month, 90 seconds each) is $46.00. If you make them with Standard, one 90-second video is $92.00, and 4 per month is $368.00. It’s 8 times more for the same number of videos.
You don’t need to make every cut with the highest-end model. Use the expensive model for the few cuts where faces or products are visible, and Lite for movement and backgrounds. Mixing them is realistic.
Fixed-cost vs. usage-based: How to choose
A fixed-cost model requires you to pay the full amount even if you don’t use it. Google’s Flow uses a credit system; according to official help, consumption for Veo 3.1 Lite is 10 credits per generation (4, 6, or 8 seconds), 20 for Fast, and 100 for Quality (8 seconds). Credits are granted as 200 per month for Google AI Plus, 1,000 for Pro, 10,000 for Ultra, and 25,000 for the next tier. I could not retrieve the monthly price in Japanese yen on September 17, but as of September 9, the Pro plan was listed at 2,900 yen per month.
With a 40% adoption rate, one 90-second short (8 seconds x 12 cuts adopted = 30 generations) requires 300 credits on Lite. If you make one per week, you need 1,200 credits, so 1,000 credits per month is not enough. If you think, ‘I can make videos for 2,900 yen a month,’ you will run out by the fourth video.
If you can predict your volume, choose a fixed-cost model; if not, choose usage-based. It is safer to use a usage-based model for the first 1-2 months to measure your generation count, then switch to a monthly plan once things stabilize. If you decide before measuring, you will usually choose a plan that is too large.
Note that I cannot confirm the presence of watermarks or the conditions for commercial use across all services, so I will not make a definitive statement. Please check the terms of service for each company yourself before signing a contract.
If you are posting to YouTube
When costs go down, you might think, ‘Can’t I win if I mass-produce?’ but there are two brakes. One is AI disclosure. YouTube’s help requires disclosure when you generate or alter content that looks realistic in a meaningful way using AI (by selecting ‘Yes’ in the Studio attribute field. It also states that disclosure does not affect monetization eligibility). The other is the ‘Inauthentic Content’ policy, which replaced ‘Repetitive Content’ on July 15, 2025. The gist is that AI-generated content made with templates lacking originality and without the creator’s perspective is not suitable for monetization. If you can make one for $0.50, anyone can make the same thing for $0.50.
A detailed table is available on my blog
I have compiled a table with the calculation process for per-second costs, 10-second clips, 60-second and 90-second shorts, Flow credit consumption, and Runway monthly plans. It is free.
What you should learn is not the price, but the formula to calculate your own cost: ‘Cost per second à Adoption rate à Duration.’ Once you have that, you can compare any service that comes out next month using the same method.
See you in another article.
Yuto
*Prices and specifications are based on information confirmed on each company’s official website as of September 17, 2026 (only the yen display for Google AI plans is as of September 9). Please be sure to check the official website and terms of service before signing a contract.
