Sora 2 and the Videos API shut down on 24 September 2026: the migration and the real cost per second

Summary. OpenAI's deprecations page lists six entries with a shutdown date of 2026-09-24: the Videos API itself, sora-2, sora-2-pro, and the three dated snapshots sora-2-2025-10-06, sora-2-2025-12-08 and sora-2-pro-2025-10-06. Developers were notified on 24 March 2026, which gives six months. The detail that matters more than the date sits in the last column of that table: the recommended replacement for every one of the six rows is a dash. OpenAI is not routing this traffic anywhere. Sora 2 was $0.10 per second of generated video and Sora 2 Pro ran $0.30, $0.50 and $0.70 per second depending on output resolution. The nearest replacements price very differently. Google's Veo 3.1 Lite is $0.05 per second at 720p and $0.08 at 1080p, Veo 3.1 Standard is $0.40, and Runway's Gen-4.5 is 12 credits per second at $0.01 per credit, so $0.12. For a team generating a thousand eight-second clips a month, that spread runs from $640 to $3,200. This is a pricing decision disguised as a deprecation notice.

What OpenAI actually published

The entry is dated 2026-03-24 and titled "Sora 2 video generation models and Videos API". The wording is short: "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." (OpenAI deprecations)

The table underneath lists six shutdowns, all on the same day.

Shutdown date	Model or system	Recommended replacement
2026-09-24	Videos API (v1/videos)	none listed
2026-09-24	sora-2	none listed
2026-09-24	sora-2-pro	none listed
2026-09-24	sora-2-2025-10-06	none listed
2026-09-24	sora-2-2025-12-08	none listed
2026-09-24	sora-2-pro-2025-10-06	none listed

Compare that with how OpenAI handles a normal deprecation. On the same page, the fine-tuned model retirements dated 23 October 2026 each carry a named target: ft-gpt-4.1-nano-2025-04-14 points at gpt-5.4-nano, ft-babbage-002 and ft-davinci-002 both point at gpt-5.4-mini. The Assistants API had a documented migration path to Responses. Video does not. On the model card, Sora 2 Pro is already tagged "Legacy". (Sora 2 Pro model card)

The practical reading: this is not a version bump you can absorb with a model string change in an environment variable. The whole v1/videos endpoint goes away, so the job-submission and polling shape of your integration goes with it.

Why the endpoint is going, in OpenAI's own words

The consumer Sora app was shut down first. Forbes reported the decision on 26 March 2026, two days after the API deprecation notice, framing it as the end of a product that had run about six months and had a Disney partnership announced in December. (Forbes)

The clearest statement of the underlying problem came from inside OpenAI, months before either shutdown. Writing on X in October, Bill Peebles, Head of Sora at OpenAI, said: "We are launching the ability to buy extra gens in Sora today. We are doing this for two main reasons: first, we have been quite amazed by how much our power users want to use Sora, and the economics are currently completely unsustainable. We thought 30 free gens/day would be more than enough, but clearly we were wrong."

Forbes lists the pressures around the decision as limited profitability, content-rights and deepfake exposure, and a compute-allocation squeeze, and is careful to label those as reasons that have been "floated around" rather than OpenAI's stated rationale. The same piece cites Wall Street Journal reporting of a $12 billion loss at OpenAI in Q3 2025. Take the causal story as unconfirmed. Take the operational fact as settled: the endpoint has a date on it.

There is a lesson for anyone building on a single vendor's generative media endpoint. Video generation is the most compute-hungry inference workload most teams will ever call, and unit economics that do not work for the vendor eventually stop working for you. The real cost is usually the migration, not the code.

What each replacement costs per second

Every provider here bills per second of generated video, which makes a like-for-like table possible. All figures below are the vendors' own published rates read on 4 August 2026.

Model and tier	Provider	Price per second of video	8-second clip
Sora 2 (720x1280 / 1280x720)	OpenAI, removed 24 Sep 2026	$0.10	$0.80
Sora 2 Pro (720x1280 / 1280x720)	OpenAI, removed 24 Sep 2026	$0.30	$2.40
Sora 2 Pro (1080x1920 / 1920x1080)	OpenAI, removed 24 Sep 2026	$0.70	$5.60
Veo 3.1 Lite (720p, with audio)	Google Gemini API	$0.05	$0.40
Veo 3.1 Lite (1080p, with audio)	Google Gemini API	$0.08	$0.64
Veo 3.1 Fast (1080p, with audio)	Google Gemini API	$0.12	$0.96
Veo 3.1 Standard (720p and 1080p)	Google Gemini API	$0.40	$3.20
Veo 3.1 Standard (4K)	Google Gemini API	$0.60	$4.80
Gen-4.5	Runway	$0.12 (12 credits)	$0.96
Gen-4 Turbo	Runway	$0.05 (5 credits)	$0.40

Sources: the Sora 2 Pro model card, the Gemini Developer API pricing page and Runway's API pricing guide. Runway bills in credits at $0.01 each, so credit rates convert directly to dollars.

Two things fall out of that table immediately.

The first is that the widely repeated figure of $0.75 per second for Veo is wrong. Google's own pricing page puts Veo 3.1 Standard video with audio at $0.40 per second for 720p and 1080p and $0.60 for 4K. If you have been budgeting from a summary article rather than the pricing page, your forecast is roughly double what Google will actually charge.

The second is that Sora 2 at $0.10 per second was not the cheap option it is remembered as. Veo 3.1 Lite undercuts it at both 720p and 1080p, and Runway's Gen-4 Turbo matches it. The expensive tier is what hurts: teams standardised on Sora 2 Pro at 1080p were paying $0.70 per second, and the closest quality-tier replacement, Veo 3.1 Standard, is $0.40. For those teams the shutdown is a cost cut.

One group will feel the change more sharply than the table suggests. OpenAI's pricing page carries a separate Batch rate, and sora-2 at 720p was listed there at $0.05 per second against the $0.10 standard rate. Teams that queued generation through Batch were paying half. There is no equivalent half-price batch lane on the per-second Veo tiers, so a Batch-heavy pipeline moving to Veo 3.1 Lite at $0.05 lands at parity rather than at a saving, and a move to Veo 3.1 Fast at $0.12 more than doubles the unit cost. Check which lane your jobs actually ran on before you present a cost forecast.

Monthly spend at three volumes

Eight-second clips, 1080p where the tier supports it, no retries counted.

Monthly clips	Sora 2 Pro (1080p, gone)	Veo 3.1 Standard	Veo 3.1 Fast	Veo 3.1 Lite	Runway Gen-4.5
250	$1,400	$800	$240	$160	$240
1,000	$5,600	$3,200	$960	$640	$960
5,000	$28,000	$16,000	$4,800	$3,200	$4,800

Retries are the line item most teams forget. Google states the rule plainly on its pricing page: "In some cases, an audio processing issue may prevent a video from being generated. You will only be charged if your video is successfully generated." Runway's documentation notes that generations routed through a Model Router are billed at the standard rate of whichever model the router picks, and that the response metadata reports the model used and the realised cost in credits, so the bill is attributable per call rather than per batch.

The four migration targets, and what each one costs you in engineering time

Google Veo 3.1 on the Gemini API

The closest thing to feature parity. Veo 3.1 comes in three variants on the Gemini API: veo-3.1-generate-preview, veo-3.1-fast-generate-preview and veo-3.1-lite-generate-preview. All three are paid tier only, with no free tier. Google labels them preview models and warns that preview models may change before becoming stable and carry more restrictive rate limits, which is a real risk to weigh when you are migrating off one deprecation and onto a preview.

The Lite variant is the volume play and carries two limitations you must design around: it does not support 4K outputs and it does not support video extension. Its documented limits are 1,024 tokens of text input and one output video per call, with text and image accepted as input and video with audio produced as output. (Veo 3.1 Lite Preview)

One trap worth naming: the older veo-3.0-generate-001 and veo-3.0-fast-generate-001 models, and veo-2.0-generate-001, were themselves deprecated with a Gemini API shutdown date of 30 June 2026. Migrating from Sora 2 to Veo 3 rather than Veo 3.1 would land you on an endpoint that has already gone. Read the version number carefully.

One operational difference will bite any pipeline that treats the provider as its archive. Google's video generation documentation states: "Generated videos are stored on the server for 2 days, after which they are removed." If your current workflow retrieves a Sora artefact lazily, or re-fetches an old job by ID, you need object storage of your own before cutover, not after. The same page also warns that "the higher the resolution, the higher the latency will be", which matters if a user is waiting on the render.

Engineering cost: moderate. You are replacing an OpenAI job-submission and polling loop with the Gemini API's video generation flow, and reworking prompt structure. Budget a sprint for a single-surface integration, more if you generate from image references.

Runway

Runway's proposition is different. Rather than one house model, its API fronts a catalogue: its own gen4.5, gen4_turbo, aleph2 and act_two, plus resold veo3.1, seedance2 and gemini_omni_flash. Everything is priced in credits at $0.01 each, so the whole catalogue sits behind one billing relationship.

Runway model	Credits per second	Effective $ per second
gen4.5	12	$0.12
gen4_turbo	5	$0.05
aleph2	28 (56 credit minimum per generation)	$0.28
veo3.1 with audio	40	$0.40
veo3.1 without audio	20	$0.20
veo3.1_fast with audio	15	$0.15
seedance2 at 1080p	40	$0.40
seedance2_mini at 480p/720p	16 (64 credit minimum)	$0.16

Note that Runway's resold veo3.1 with audio at $0.40 per second matches Google's own Standard rate exactly, so the aggregation does not carry a markup at that tier. It does carry an important structural feature: Model Routers. Runway documents the selection in two steps: "First it narrows the catalog to the models that could satisfy the request - those that you've enabled, support the capabilities required by your request, and would stay under any price cap you've set. Then it picks the best of the remaining models according to your preference for cost, latency, or quality." The price cap is the part worth noting, because it turns a per-second rate into a governed budget rather than an open tap, and "the JSON response includes which model ran and what it cost". Runway's pricing page ends its deprecation footnote with a pitch to "never need to manage deprecations again", which is a fair summary of what a router buys you. It also demonstrates the problem in the same breath. That footnote reads: "Veo 3 is deprecated and will be officially sunset on August 4th, 2026" — that is today. The video-model layer is churning fast enough that two named models on two different platforms retired within eight weeks of each other.

Engineering cost: low to moderate. One API surface, one billing relationship, and the router absorbs the next retirement. The trade is a dependency on an intermediary.

Gemini Omni Flash

Google's video generation and editing model bills on tokens rather than seconds: $1.50 per million input tokens for text, image, video or audio, and on the output side $9.00 per million text tokens and $17.50 per million video tokens on the Gemini API paid tier. That makes direct per-second comparison impossible without measuring your own token consumption, which is exactly the sort of forecasting work worth doing before you commit. If your pipeline is edit-heavy rather than generate-from-scratch, model it properly rather than assuming the per-second models are cheaper. We have covered the Gemini Omni Flash video cost model for marketing teams separately.

Staying multi-model behind a gateway

The structural answer to a vendor removing an endpoint with no successor is to stop calling vendors directly. The same routing and fallback pattern teams already run for text models applies here: a gateway holds the credentials, normalises the request shape, and lets you repoint a workload without touching application code. Our comparison of LiteLLM, Cloudflare, Kong and Bifrost as AI gateways covers the trade-offs, and the hybrid routing and API spend decision framework sets out when routing pays for its own complexity.

A migration plan with dates

Six weeks remain between today, 4 August 2026, and the 24 September 2026 shutdown. That is enough time if you start now and short if you wait for a September freeze window.

1. Inventory the calls. Grep for v1/videos, sora-2 and sora-2-pro across application code, background workers, notebooks, internal tools and any scheduled job. The dated snapshots are the ones that hide: sora-2-2025-10-06, sora-2-2025-12-08 and sora-2-pro-2025-10-06 are pinned in exactly the places a team pinned them for output stability and then forgot.
2. Classify by tier, not by model. Work out how much of your volume was on Sora 2 at $0.10 and how much on Sora 2 Pro at $0.30 or above. Those two groups migrate to different targets. The cheap tier goes to Veo 3.1 Lite or Gen-4 Turbo. The quality tier goes to Veo 3.1 Standard or Gen-4.5.
3. Rebuild the job loop, not just the model string. The Videos API is asynchronous: submit, poll, retrieve. Whatever you move to has its own job lifecycle, its own terminal states and its own artefact retrieval. This is where the sprint goes.
4. Re-run your prompt library. Prompts tuned against Sora 2 do not transfer cleanly. Build a small golden set of twenty to thirty prompts that represent your real output mix, generate against each candidate, and have a human score them before you commit spend.
5. Check the rate-limit ceiling. Sora 2 Pro's limits ran from 10 requests per minute at Tier 1 up to 150 at Tier 5, with no free tier at all. If you were operating near the ceiling on OpenAI, confirm the equivalent headroom on the new provider before cutover, not after.
6. Cut over with a kill switch. Keep both paths live behind a flag until the last week of September. The endpoint disappears on a fixed date, so there is no partial-degradation phase to detect. It works, and then it returns errors.

India-specific considerations

For teams in India generating marketing or product video, three points change the calculation.

Currency exposure is the first. Every provider on this list bills in US dollars, so a video pipeline is a dollar-denominated operating cost sitting inside a rupee P&L. At the volumes in the table above, moving from the Sora 2 Pro tier to Veo 3.1 Lite is the difference between a $28,000 and a $3,200 monthly line at 5,000 clips. That is a large enough gap to justify a proper evaluation rather than a default choice.

Data handling is the second. If your generation prompts or reference images contain customer photographs, employee likenesses or anything else that qualifies as personal data, the pipeline falls inside the Digital Personal Data Protection Act 2023. The relevant questions are the ordinary ones: what is the lawful basis for putting that image into a third-party model, is it captured in your notice and consent flow, and what does the provider's retention policy say. Google's pricing page carries a per-model "used to improve our products" column, set to No on the paid tier for the Veo models and Yes on the free tier, which is a material distinction for anyone processing personal data.

Likeness and rights exposure is the third. Content-rights and deepfake concerns are among the pressures Forbes cites around the Sora decision. If your pipeline generates people, the governance question does not go away when you change vendors.

FAQ

When exactly does the Sora 2 API stop working?

OpenAI's deprecations page gives a single shutdown date of 24 September 2026 for all six affected entries: the Videos API, sora-2, sora-2-pro, and the snapshots sora-2-2025-10-06, sora-2-2025-12-08 and sora-2-pro-2025-10-06. Developers were notified on 24 March 2026, six months ahead of removal.

Has OpenAI named a replacement model?

No. The recommended replacement column is empty for all six entries, unlike other deprecations on the same page where retiring fine-tuned models point at named successors such as gpt-5.4-nano and gpt-5.4-mini. Sora 2 Pro's model card is already labelled Legacy. There is no OpenAI video endpoint to migrate to.

What did Sora 2 cost per second?

Sora 2 was $0.10 per second of generated video at 720x1280 portrait or 1280x720 landscape. Sora 2 Pro was $0.30 per second at those resolutions, $0.50 at 1024x1792 or 1792x1024, and $0.70 at 1080x1920 or 1920x1080. Both are removed on 24 September 2026.

Which replacement is cheapest?

Veo 3.1 Lite on the Gemini API is the lowest published rate at $0.05 per second for 720p and $0.08 for 1080p, both with audio. Runway's Gen-4 Turbo matches the 720p figure at 5 credits per second, which is $0.05. Veo 3.1 Lite does not support 4K output or video extension.

Is Veo 3.1 really $0.75 per second?

No. Google's Gemini Developer API pricing page lists Veo 3.1 Standard video with audio at $0.40 per second for 720p and 1080p and $0.60 per second for 4K. Veo 3.1 Fast is $0.10 at 720p, $0.12 at 1080p and $0.30 at 4K. The higher figure circulating in summaries does not match Google's own page.

Should we move to Veo 3 or Veo 3.1?

Veo 3.1. The Veo 3 models veo-3.0-generate-001 and veo-3.0-fast-generate-001, and Veo 2's veo-2.0-generate-001, were deprecated with a Gemini API shutdown date of 30 June 2026. Runway separately lists its resold veo3 as sunset on 4 August 2026. Migrating to Veo 3 means migrating twice.

How long should the migration take?

Plan a sprint per generation surface. Replacing the model string is trivial; rebuilding the asynchronous submit-poll-retrieve loop, re-tuning a prompt library against a new model, and re-validating rate-limit headroom is the actual work. With six weeks left before 24 September 2026, starting in August leaves room for a scored quality comparison.

Does a gateway solve this?

Partly. A gateway or model router holds credentials and normalises request shapes so a workload can be repointed without application changes, which shortens the next migration. It does not remove the prompt-tuning work or the quality evaluation, because those are model-specific. Runway's Model Routers and independent AI gateways both address the routing layer only.

How eCorpIT can help

eCorpIT is a Gurugram-based technology consultancy, founded in 2021, with senior engineering teams working across AI integration, cloud and application development. We are CMMI Level 5, ISO 27001:2022 certified and MSME registered. If you are running generative video in production and need the migration scoped, costed against your real clip volumes and delivered before the 24 September 2026 cutoff, we can run the inventory, build the model comparison against your own prompt set, and implement the cutover behind a feature flag. Our teams also cover the LLM migration and cost optimisation work that usually sits alongside it. Talk to us at /contact-us/.

References

1. Deprecations — OpenAI API: the 2026-03-24 entry listing all six Sora 2 and Videos API shutdowns on 2026-09-24 with no recommended replacement.
2. Sora 2 Pro Model — OpenAI API: per-second video pricing of $0.30, $0.50 and $0.70 by resolution, the Sora 2 comparison at $0.10, the Legacy tag and the tier rate-limit table.
3. Gemini Developer API pricing — Google AI for Developers: Veo 3.1 Standard, Fast and Lite per-second rates, Veo 3 and Veo 2 shutdown warnings dated 30 June 2026, Gemini Omni Flash token pricing.
4. Veo 3.1 Lite Preview — Gemini API: the 4K and extension limitations, input and output modalities, and 1,024-token text input limit.
5. API Pricing and Costs — Runway Dev: the full credits-per-second table, the $0.01 credit price, Model Router billing behaviour and the Veo 3 sunset footnote dated 4 August 2026.
6. Video generation in the Gemini API — Google AI for Developers: the resolution and extension feature documentation referenced by the Veo 3.1 model pages.
7. OpenAI Shuts Down AI Video App Sora — Forbes: the 26 March 2026 report, the Bill Peebles quotation, the Q3 2025 loss figure attributed to the Wall Street Journal, and the content-rights and compute pressures.
8. Veo 3.1 — Gemini API model page: the veo-3.1-generate-preview model identifier and preview status.
9. Models — OpenAI API: the current model catalogue showing which OpenAI endpoints remain after the video removal.
10. Pricing — OpenAI API: the platform pricing page referenced from the Sora 2 Pro model card.
11. Runway Models guide: the model catalogue behind the Runway API, including resold third-party models.
12. Model Routers — Runway Dev: how router selection and per-call cost attribution work.

Last updated: 4 August 2026.