AI Video Models: A September 2026 Snapshot

We have argued elsewhere that best-of rankings for AI video models are obsolete the day they publish. That argument still holds, so this is not a ranking. It is a dated snapshot: what the landscape looks like in September 2026, what changed, and what a working pipeline should do about it.

The most important item is not a capability. It is a deadline.

Sora 2 shuts down on 24 September 2026

From OpenAI's own deprecations page: "The Sora 2 video generation models and Videos API are deprecated and will shut down on September 24, 2026."

That covers sora-2, sora-2-pro, sora-2-2025-10-06, sora-2-2025-12-08 and sora-2-pro-2025-10-06. Developers were notified on 24 March 2026, and the retirement was announced publicly the following day. The Sora web and app experiences were discontinued on 26 April 2026, and the Disney partnership collapsed in the same period.

From notice to shutdown

Sora 2 retirement, key dates in 2026, spaced to scale

From notice to shutdown

Date	Event
24 March 2026	Developers notified
26 April 2026	Sora web and app experiences discontinued
24 September 2026	Sora 2 video generation models and Videos API shut down; no replacement named

No replacement model has been named. This is not a migration path, it is a removal. If any part of your production stack, any automation, any client deliverable schedule depends on the Sora API, you have days rather than months, and the destination is another vendor entirely.

Worth pausing on what this says about platform risk in general. A frontier video model from the best-capitalised lab in the sector went from launch to shutdown inside a year, taking an entire API surface with it. Any pipeline built on a single vendor's video endpoint carries that risk structurally, regardless of which vendor.

Who is at the top, and why the number is nearly useless

Two public leaderboards carry weight right now, and they do not agree, because they cannot.

• On arena.ai's text-to-video board, gemini-omni-1.1-flash sits at roughly 1515 Elo.
• On Artificial Analysis's text-to-video board, Wan 3.0 leads at roughly 1242, with 30-second clip generation.

Those two numbers are not comparable. Elo is relative to the pool of models being compared and the population of voters, not an absolute scale. A 1515 on one board and a 1242 on another tell you the ordering inside each board and nothing across them. Anyone presenting them in a single table is producing a chart, not a comparison.

The more instructive data point is what happened to Runway Gen-4.5. In December 2025 it was ranked first at 1247 Elo. It is now outside the top ten on both boards. Nine months, first to unranked, with no drop in the model's actual quality. That is the argument against rankings in a single example.

What actually distinguishes models this month

Ordering aside, there are real structural differences worth knowing, because these are the things that change what you can build rather than what you score.

• Clip length. Wan 3.0 generating 30-second clips is a workflow change, not an increment. Most assembly complexity in AI video comes from stitching short clips, and longer native generations remove cuts you were previously forced into.
• Licensing. LTX-2 is open-weight and free to use commercially below $10M ARR. For a studio, open weights mean the model cannot be deprecated out from under you, which after the Sora 2 item above is not an abstract benefit.
• Native resolution and frame rate. LTX-2 generates natively at 4K and 50fps. Native beats upscaled, and 50fps matters for anything that will be slowed down or broadcast in PAL territories.
• Release status. Veo 3.1 remains in Preview. Preview means terms, pricing and availability can move, and it is a poor foundation for a committed client schedule.

The method that outlives the snapshot

Everything above expires. The way you choose does not. Four questions, in order of how much they constrain the decision.

• What is the deprecation risk? Ask it first now, not last. An open-weight model you can run yourself has a different risk profile from a Preview endpoint at a lab that has already killed one video product.
• Does it do the specific thing this project needs? Reference conditioning, first and last frame control, extension, native vertical, lip sync. A model that leads a general board and lacks the one control your project depends on is the wrong model.
• What does a usable shot cost? Not what a generation costs. The number that matters is cost per accepted shot, which is generation cost multiplied by your re-roll rate, and the re-roll rate is only knowable from your own tests.
• Can the output be graded? Bit depth and colour handling decide how much of the look you have to nail in generation versus fix in post. This is invisible on every leaderboard.

What we would do this week

• Audit for Sora 2 today. Grep your automations, scripts, n8n or Make workflows and client SOWs for sora. This is a ten-minute check with an eleven-day deadline.
• Stand up one open-weight fallback. Not as a primary, as insurance. The Sora 2 shutdown is the second time in a year that a production stack has needed one.
• Test long-clip generation against your own assembly. If 30-second native clips remove two cuts per scene from your workflow, that is a real time saving you can measure this month.
• Stop quoting cross-board Elo internally. Pick one board, state its date, and treat it as a shortlist generator rather than a verdict.

The short version

September 2026 has one urgent fact and one durable lesson. The urgent fact is that Sora 2 and the Videos API are gone on 24 September with nothing named to replace them. The durable lesson is that a model's leaderboard position tells you less about your project than its licence, its controls and its deprecation risk, and those three are knowable without waiting for anyone's ranking to update.