The Last Frame: OpenAI's Sora 2 API Goes Dark Today With No Successor

Today, September 24, 2026, is the day OpenAI’s video generation business officially ends. The Videos API and every sora-2 model alias and snapshot are removed from the OpenAI API, closing out a wind-down that began with a developer notice on March 24 and escalated through the shutdown of the standalone Sora web and app experiences on April 26. What makes today’s deadline unusual — and what has developers on forums and social media frustrated — is what’s missing from the official deprecations table: a recommended replacement. The column is simply blank.

A Six-Month Wind-Down, Not a Sudden Death

To be fair to OpenAI, nobody can claim they were ambushed. The company notified developers a full six months ago, on March 24, 2026, that the Videos API and the Sora 2 model family would be retired on September 24. The consumer-facing Sora app — launched with enormous fanfare as OpenAI’s big push into social video — lasted roughly six months before being discontinued on April 26, 2026. Since then, Sora 2 remained reachable only through ChatGPT subscriptions and select API partners, a kind of hospice period that ended today.

The timeline tells a story of a product line being unwound in stages rather than a technology being upgraded:

• December 2024: Original Sora launches on the web after the headline-grabbing February 2024 preview.
• Late 2025: Sora 2 launches, including the standalone app with a Cameo feature.
• March 24, 2026: OpenAI notifies developers of the API deprecation; the deprecations table lists no successor.
• April 26, 2026: Sora web and app experiences discontinued.
• September 24, 2026: Videos API and all sora-2 aliases and snapshots removed.

Automation platforms have been scrambling on their own schedules. Make, for instance, retired its OpenAI Sora and video modules ahead of today’s deadline, warning users to rebuild workflows before the cutoff.

Why the Replacement Column Being Blank Matters

Model deprecations at frontier labs normally follow a predictable pattern: an old model is retired and a newer one is slotted into the “recommended replacement” field. GPT-4 gives way to GPT-4.1, then to GPT-5-family models. Even when developers grumble about pricing or behavior changes, there is a path forward within the same API.

Sora 2 is getting no such handoff. OpenAI’s own documentation directs developers to nothing — no Sora 3, no video-capable GPT-6 endpoint, no partner integration blessed as the official successor. As one summary of the deprecations table put it bluntly: the shutdown date is confirmed, and the replacement column is blank.

The community’s reaction has been more than grumbling. A thread on the OpenAI developer forum titled “Release Sora 2 open weights before shutdown” has become a rallying point, asking the company to open-source the model rather than let it vanish entirely. The petitioners’ argument is straightforward: the model already exists, its serving infrastructure is being decommissioned anyway, and releasing the weights would cost OpenAI little while preserving years of community tooling, ComfyUI integrations, and migration investment built around Sora 2’s API surface. OpenAI has not, as of this writing, signaled any intention to do so — the company that once championed “open” in its name will let the weights die with the API.

What Happened: A Market That Moved Faster Than OpenAI

Post-mortem analyses of the Sora app’s collapse converge on competitive pressure. When OpenAI’s original Sora preview stunned the world in February 2024, video generation was a two-horse race and OpenAI looked a generation ahead. Two years later, the field is crowded and OpenAI is not leading it.

Google’s Veo line — now at Veo 3.1 — is widely regarded as the quality benchmark, with strong physics realism and tight integration into the Gemini ecosystem. Kuaishou’s Kling dominates motion realism and physics among independent creators. ByteDance’s Seedance models compress generation time dramatically and piggyback on ByteDance’s content-platform distribution. Runway’s Gen-4.5 holds the “cinematic” positioning with professional colorists and film crews. Against that lineup, Sora 2’s app struggled to differentiate, and the API business — expensive to serve, with video inference costs far above text — became a cost center rather than a strategic asset.

There is also a portfolio logic at work. OpenAI’s 2026 narrative has been consolidation: GPT-6 Sol and Luna launched last week as cheaper, focused frontier models; agentic surfaces like the new ChatGPT voice plugins are where the company is investing. Every dollar spent keeping a losing video product alive is a dollar not spent on the agent race against Anthropic and Google. Killing Sora is the kind of decision a company makes when it has decided where its actual war is.

What Developers Should Do Now

If you have production workflows that still call the Videos API, they are failing as of today. The realistic options, in rough order of migration effort:

1. Veo 3.1 (Google) — closest to a drop-in quality upgrade, with documented API pricing per second and the deepest ecosystem support.
2. Runway Gen-4.5 — strongest for cinematic and professional post-production pipelines.
3. Kling — the pick when motion realism matters most.
4. Seedance (ByteDance) — fastest generation times, especially relevant if you were using Sora for high-volume, short-form content.
5. Self-hosted open models — for teams with GPUs and privacy requirements, the open-weight video model ecosystem has matured enough that Sora 2’s exit is survivable without any closed API at all.

The Bigger Picture

The Sora exit is worth marking for what it says about the AI industry’s shape in late 2026. First, video generation is commoditizing faster than anyone expected — a frontier capability in 2024 is now a competitive bloodbath where even OpenAI can’t find a defensible position. Second, OpenAI is willing to kill a flagship brand outright rather than maintain a zombie product, which tells you something about internal resource discipline ahead of what analysts widely expect to be a listing process. And third, the refusal to release the weights — even as a goodwill gesture to the developer community that built on the API — is a data point in the ongoing debate about what “open” means in the GPT-6 era.

Sora’s first preview frame was one of the most-shared images in AI history. Today the last API call returns an error. Between those two moments lies the full arc of the generative video market’s first chapter — and a reminder that in this industry, being first counts for nothing if you aren’t still the best.