Direct answer: on 24 September 2026, OpenAI's Videos API and the Sora 2 video models shut down under the official deprecation notice, and the "recommended replacement" column is empty — there is no one-to-one replacement API. The lesson is not "one fewer video model". It is that procurement of enterprise AI must treat vendor shutdown, notice period, asset export and migration cost as part of the contract, right next to price. This guide uses OpenAI and Google official lifecycle pages to give a 2026–2027 shutdown calendar, the two vendors' notice commitments, a recomputable exit-cost formula, and eight contract clauses.
It is written for technology, procurement, legal and business owners who are buying AI capability or already run it in production. Data cut-off: 31 October 2026; only vendor official pages are used. It does not predict which vendor will shut down next, does not benchmark model quality, and is not legal advice or a quote. Except for official list prices, every amount below is an editorial example assumption.
Shutdown is part of a product lifecycle, not an accident. The real risk is "we did not know early + assets could not be exported + no alternative". Put notice period, export window, replacement path and termination assistance into the contract — cheaper than comparing prices after the shutdown.
What exactly shut down in September 2026?
OpenAI announced on 2026-03-24 and shut down on 2026-09-24: the Videos API, sora-2, sora-2-pro and their snapshots. No one-to-one replacement was given; download links last at most 1 hour, and batch outputs 24 hours.
| Item | Official fact | Date / window |
|---|---|---|
| Deprecation notice | Developers using the Videos API and Sora 2 aliases/snapshots were notified | 2026-03-24 |
| What shut down | Videos API, sora-2, sora-2-pro, plus the 2025-10-06 / 2025-12-08 snapshots | Unavailable from 2026-09-24 |
| Recommended replacement | The official column shows ---: no one-to-one replacement API | — |
| Generated videos | Download links valid for at most 1 hour; batch outputs for up to 24 hours | Export before expiry |
| Content restrictions | Real people (including public figures), copyrighted characters and copyrighted music are rejected; content must suit under-18 audiences | Affects advertising and brand films |
| Procurement meaning | After shutdown you cannot call it again — only migrate or switch vendor; existing assets must be exported first | — |
Sources: OpenAI "Deprecations" and "Video generation with Sora". "No replacement" reflects the official page, not the absence of other video models on the market.
Shutdown is not a one-off: the official 2026-11 to 2027-04 calendar
OpenAI has eight more batches between 2026-11-30 and 2027-04-01; Google's Gemini 2.5 series already retired on 2026-10-20, and Veo 3.1 retires no earlier than 2026-11-17. These are published dates, not forecasts.
| Date | Vendor | What is retiring | Procurement implication |
|---|---|---|---|
| 2026-10-20 | gemini-2.5-pro, gemini-2.5-flash, gemini-2.5-flash-lite | Already retired; re-check any quote that still lists them | |
| 2026-11-17 or later | veo-3.1-generate-001, veo-3.1-fast-generate-001 | At least 12 months after release; dates only move later | |
| 2026-11-30 | OpenAI | Reusable prompt objects API; Evals dashboard/API; Agent Builder | Evaluation and builder tooling goes first — export eval sets early |
| 2026-12-01 | OpenAI | gpt-image-1-mini, gpt-image-1.5, chatgpt-image-latest | Image pipelines need a new model and a regression run |
| 2026-12-11 | OpenAI | gpt-5 / mini / nano / pro; o3 / o3-pro | Apps pinned to old snapshots need migration within two months |
| 2026-12-13 | gemini-live-2.5-flash-native-audio | Realtime voice/customer-service paths need alternatives | |
| 2027-01-06 | OpenAI | tts-1, tts-1-hd, gpt-4o-mini-tts; existing customers can no longer create fine-tuning jobs | Voice-over and fine-tuning pipelines hit the same date |
| 2027-01-14 | OpenAI | gpt-4o, gpt-4o-mini, gpt-4.1, gpt-4.1-mini (and their fine-tuned versions) | The widest impact — many production apps sit here |
| 2027-01-20 | OpenAI | gpt-realtime, gpt-audio, gpt-4o-audio, gpt-realtime-mini, gpt-audio-mini and related | Voice agents and live captions must re-test latency and audio |
| 2027-02-26 | OpenAI | whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-4o-transcribe-diarize, /v1/audio/translations | Transcription and translation APIs change together |
| 2027-03-15 | gemini-2.5-flash-image | Another image-pipeline migration in Q1 2027 | |
| 2027-04-01 | OpenAI / Google | OpenAI gpt-5.1, gpt-5.3-codex, gpt-5.4-nano; Google text-embedding-004 / -005, multilingual and multimodal embeddings | Coding assistants and vector stores regress together |
Sources: OpenAI "Deprecations" and Google Cloud "Model versions and lifecycle". Only entries with explicit dates are listed; "or later" is the vendor's wording, not an editorial forecast.
How different are the notice periods?
OpenAI commits to at least 6 months for generally available models, 3 months for specialized variants, and as little as about 2 weeks for preview models. Google's 12-month list commits to at least 12 months after release and dates only move later; short-term models get at least 45 days.
| Category | OpenAI public commitment | Google public commitment |
|---|---|---|
| Flagship / GA | At least 6 months | 12-month list: at least 12 months after release |
| Specialized variants | At least 3 months (chat, codex, deep research, etc.) | Listed per model |
| Preview / short-term | Can be as short as about 2 weeks | Short-term availability models: at least 45 days to migrate |
| Can it be earlier? | Yes for safety or compliance, with notice as early as reasonably possible | Dates on the list are not brought forward |
| Procurement meaning | Do not put production on preview models; budget migration on 6 months | Annual contracts are feasible, but short-term models need separate review |
Both commitments live in documentation, not in your contract; sign the "no shorter than the vendor's public notice period" obligation into the agreement.
Why does a shutdown become a budget line?
Because the cost is not "buy another API". It is integration rework, asset export, regression and acceptance, dual-running overlap, compliance re-review and business interruption. Official download windows (1 hour / 24 hours) also compress the export.
| Cost line | How to meter | Example parameter (assumption) | Example amount (assumption) |
|---|---|---|---|
| Integration rework | Interface/auth/params/retry days × internal day rate | 6 person-days | CNY 7,200 |
| Asset and prompt export | Export, verify, backfill days | 2 person-days | CNY 2,400 |
| Regression and acceptance | Golden-set rerun and human rating days | 4 person-days | CNY 4,800 |
| Compliance re-review | Content policy, copyright, labelling checks | 2 person-days | CNY 2,400 |
| Dual-running overlap | Duplicate usage fees until the old system is off | 1 month × CNY 3,000 baseline | CNY 3,000 |
| Business interruption / reshoot | Project-specific, commonly omitted | Not included | — |
| Total (example) | 14 person-days + 1 month overlap | Day rate assumed at CNY 1,200 | From CNY 19,800 |
Formula: exit cost = (integration + export + regression + compliance) person-days × day rate + overlap cost + interruption. Person-days, day rate and monthly fee are example assumptions showing a recomputable structure; they are not a quote or any vendor's actual price.
The same logic applies to billing units. Google's pricing page lists Veo 3.1 at "USD 0.40 / 1 count" with audio at 720p/1080p and "USD 0.20 / 1 count" for video only, but the same page does not define how a count maps to seconds or clips; Veo 3.1 generates 4-, 6- or 8-second videos and up to 4 outputs per prompt. This article therefore does not convert counts into a per-video price. Before signing, run one 4-second and one 8-second job and reconcile the invoice line items — cheaper than a later dispute.
Migration / exit acceptance checklist
Acceptance answers four questions: are all assets exported, does the alternative pass on the same rubric, is the overlap period controlled, and is the old system actually retired.
| Acceptance item | Suggested definition | Evidence to keep |
|---|---|---|
| Asset inventory | Prompts, reference images/videos, character assets, outputs, eval sets registered item by item | Export list, checksums, storage location |
| Export timing | Complete inside the official window (e.g. 1 hour for video, 24 hours for batch) | Export timestamps, retry log |
| Alternative | Same tasks, same rubric, same resolution — not a vendor demo | Comparison results, failure samples |
| Cost comparison | Same billing unit; if unclear, run an invoice test first | Two invoices, conversion notes |
| Compliance | Content policy, copyright, likeness, AI labelling | Review records, asset licences |
| Overlap period | Explicit start/end, duplicate-cost cap, rollback trigger | Overlap plan, rollback drill |
| Retirement | Old keys revoked, callbacks cleaned, monitoring switched | Revocation record, monitoring screenshot |
| Retrospective | Was the shutdown detected early, was notice enough | Findings, contract amendments |
Editorial guidance, not an audit standard. High-risk workloads need legal, security and business sign-off.
Which eight clauses belong in the contract?
Notice period, replacement commitment, export SLA, price protection, model pinning, regional availability, content-policy change and termination assistance decide whether you migrate on plan or firefight.
| Clause | Suggested wording | Unacceptable |
|---|---|---|
| Notice period | No shorter than the vendor's public commitment (e.g. 6 months for GA) | Only "reasonable advance notice" |
| Replacement | Shutdown must come with equivalent capability or a defined migration path | "No replacement", as with the video API |
| Export SLA | Export window, formats, failure support | A 1-hour download link only |
| Price protection | Unit-price cap and notice period for increases | Unilateral repricing |
| Model pinning | Named snapshot versions with a minimum availability period | Automatic "latest" |
| Regional availability | Data residency, cross-border, availability-zone commitments | Region changes at will |
| Policy change | Material content-policy changes need notice and a termination right | Restrictions revealed at shutdown |
| Termination assistance | 30–90 days of transition support with a fee cap | Hard cut-off, no assistance |
Editorial guidance, not legal advice; final wording depends on applicable law and your negotiating position.
A 20-minute monthly "shutdown radar"
You do not need daily news monitoring: one monthly pass over the official deprecation pages, the model-list API and your internal dependency register turns firefighting into planned migration.
- Read the deprecation pages: OpenAI Deprecations and Google Model versions and lifecycle; record new entries and dates.
- Check the API field: OpenAI's
GET /v1/modelsreturns an optionalshutdown_dateyou can monitor directly. - Reconcile the register: every production app records its model/endpoint, owner and last regression date.
- Quantify exposure: rank dependencies expiring within six months by business impact; handle revenue-linked paths first.
- Refresh the budget: apply the six-line formula to your own day rate and monthly fee, once a quarter.
Who should read this — and who should not?
For teams already running AI in production or signing API contracts. Not for teams that only benchmark models or have no production dependency yet.
Read if: you are a technology lead, procurement/legal, platform or operations team; you are signing an API contract, putting video/voice/image capability into production, or have already been hit by a vendor shutdown. Skip if: you only run proofs of concept, only care about benchmark scores, or want this article to choose a vendor for you. Supports: shutdown risk inventory, migration budget, contract negotiation, exit drills and quarterly reviews.
Lingyi does not rank model vendors and cannot promise that any vendor will not shut down. If you need multi-model routing, fallback switching, and centralised key and audit management, see LingHarness; to fold shutdown risk into project acceptance and budgets, talk to us.
Related reading: LLM API cost, fine-tuning cost and RAG decisions, private-deployment TCO, and project stop-loss and exit acceptance; this article only covers vendor/model lifecycle shutdown and migration cost.
FAQ
Can we still use OpenAI's video API?
No. The official deprecation notice lists the Videos API, sora-2, sora-2-pro and their snapshots as shut down on 2026-09-24 with an empty "recommended replacement" column — there is no one-to-one replacement API. Video capability now means another vendor plus new integration and acceptance work.
Are videos generated before the shutdown still available?
Per the official guide, download links are valid for at most 1 hour after generation, and batch outputs for up to 24 hours. The vendor makes no long-term storage commitment, so exporting assets into your own storage before shutdown is a hard requirement.
How long is a typical shutdown notice?
OpenAI publicly commits to at least 6 months for GA models, 3 months for specialized variants, and as little as about 2 weeks for preview models, with shorter notice possible for safety or compliance. Google's 12-month list commits to at least 12 months after release and dates only move later; short-term models get at least 45 days. Documentation commitments still need to become contract terms.
How do we detect an upcoming shutdown early?
Three habits: read the official deprecation pages monthly; monitor the shutdown_date returned by OpenAI's GET /v1/models; keep an internal dependency register with an owner. About 20 minutes a month, far cheaper than a shutdown-day scramble.
How should migration cost be estimated?
Use the six-line formula: integration, asset export, regression/acceptance, dual-running overlap, compliance re-review, business interruption. Evaluation and compliance re-review are the most commonly omitted — they add no features but decide whether the migration is really finished.
Which single clause matters most?
If you can add only one: "shutdown notice no shorter than the vendor's public commitment, plus export assistance and a transition period". The video API case shows that when there is no replacement, notice and export windows are the only buffer a customer has.
Why does the billing unit belong in the contract?
Because "per count", "per credit" and "per second" are different units. Google's pricing page lists Veo 3.1 per 1 count without defining the conversion on the same page, while the model generates 4/6/8-second clips. Run different durations, reconcile the invoice, and only then treat the unit price as budgetable.
Method and limitations
This article uses only vendor official pages accessible on 31 October 2026: OpenAI's deprecation notice and video-generation guide, and Google Cloud's model lifecycle, Veo 3.1 documentation and Vertex AI pricing page. It does not use private contracts, customer invoices, internal roadmaps or third-party forecasts. Shutdown dates can be extended (both vendors reserve that right); the article records published dates and does not infer unpublished ones. Person-days, day rates and monthly fees in the exit-cost table are example assumptions, not any company's actual data and not a quote; integration complexity varies widely. The article does not evaluate model quality and is not legal advice.
Sources and basis
Data cut-off: 31 October 2026. The following are official pages; dates, models and terms follow the original text.
- OpenAI, Deprecations: developers.openai.com/api/docs/deprecations. Sora 2 / Videos API shutdown on 2026-09-24 with no replacement; 6-month GA, 3-month specialized and ~2-week preview notice commitments; remaining 2026–2027 shutdown dates; the
shutdown_datefield. - OpenAI, Video generation with Sora: platform.openai.com/docs/guides/video-generation. Shutdown note, 1-hour download links / 24-hour batch outputs, content restrictions.
- Google Cloud, Model versions and lifecycle: cloud.google.com/vertex-ai/generative-ai/docs/learn/model-versions. 12-month list, Veo 3.1 retirement no earlier than 2026-11-17, at least 45 days for short-term models, Gemini 2.5 retired 2026-10-20.
- Google Cloud, Veo 3.1 generate: cloud.google.com/vertex-ai/generative-ai/docs/models/veo/3-1-generate. 4/6/8-second videos, up to 4 outputs per prompt.
- Google Cloud, Vertex AI generative AI pricing: cloud.google.com/vertex-ai/generative-ai/pricing. Veo 3.1 / Fast / Lite price per 1 count by resolution; the page does not define a count-to-second or count-to-clip conversion.