Sora is Dead. OpenAI Pulled the Plug to Pour GPUs Into ChatGPT Images V2 - Here's the Real Reason
On April 28, 2026, OpenAI announced that Sora - its flagship text-to-video model - was being sunset. Existing customers have until July to migrate; the API is already accepting no new signups. The official phrasing was that OpenAI is "focusing on the products users love most." The real story, as anyone watching GPU spot prices already knows, is that video generation is a money-loser at current compute costs, and ChatGPT Images V2 isn't.
What Was Actually Announced
- Sora API: Closed to new signups effective immediately. Existing accounts get rate-limited access through July 31, 2026, then full shutdown.
- Sora consumer product: Folded into ChatGPT Plus/Pro as a sharply reduced "video moment" feature - 5-second clips only, queued generation, no sound.
- Customer migration: Existing video generation credits convert 1:5 into ChatGPT Images V2 credits. The implicit message: we'd rather you make 5 still images than 1 video.
- Sam Altman's tweet: "We learned a lot from Sora. The pattern of usage told us where users get the most value, and we're going where that signal points."
Translation: we burned a lot of GPU-hours on a product that didn't pay back its compute. We're not doing that anymore.
The Brutal Math of Video Generation
Here's why Sora was always going to be a loss leader at current GPU prices. Generating one minute of high-quality 1080p video with a transformer-based video model costs roughly the same compute as generating 800 to 1,500 still images of similar quality - depending on frame rate, resolution, and the complexity of the temporal coherence pass.
| Output | Approximate GPU-seconds | Price OpenAI charged | Implied $/GPU-hour |
|---|---|---|---|
| 1 ChatGPT Image (1024×1024) | ~3–6 | $0.04 (or free in Plus) | ~$24–$48 |
| 1 minute of Sora 1080p video | ~3,000–5,000 | $3–$8 (subsidized) | ~$3–$10 |
| 1 minute of Sora 4K video | ~12,000–20,000 | $15–$25 (heavily subsidized) | ~$3–$7 |
OpenAI was effectively renting out H100 time at well under cost on Sora to keep the product attractive, while ChatGPT Images V2 - with its dramatically smaller compute footprint - generates handsome margin and is bundled into a $20/month subscription that users already pay for.
Use our price comparison tool to see how compute costs trickle into per-output pricing across providers. The same dynamic plays out across every modality.
ChatGPT Images V2 - The Other Side of the Trade
ChatGPT Images V2 launched in late April 2026 to widespread acclaim (covered in our deep-dive on the design industry impact). It does several things that make it dramatically more profitable per GPU-second than Sora:
- Bundled into existing subscriptions. ChatGPT Plus and Pro users already pay $20/$200 a month. Image generation comes "for free," but the marginal compute cost per image is tiny.
- High retention. Image users come back daily - social posts, slides, mockups, thumbnails. Video users in Sora came back maybe once a week.
- Viral by nature. Every generated image gets shared, embedded, screenshotted. Free marketing.
- Lower regulatory exposure. Video deepfakes are a political and legal minefield. Stills are still risky but operationally easier to moderate.
The Industry-Wide Compute Reallocation
OpenAI is not the only lab making this trade. Every frontier AI provider is staring down the same constraint: finite GPU supply, exploding demand. When you have to choose between two products that both need more compute than you can give them, you pick the one with the better unit economics. We're seeing the same pattern everywhere:
| Lab | Compute reallocation | Strategic logic |
|---|---|---|
| OpenAI | Sora → ChatGPT Images V2 | Bundle into $20 subscription, viral distribution |
| Anthropic | Long-context Opus 4.6 → Opus 4.7 with peak throttle, then 4.7 with relaxed limits after Google's $40B investment (see our breakdown) | Capacity-starved before, breathing room now |
| Veo 3 still in research, no aggressive consumer push | TPU advantage means video is less margin-pressured - but still not the priority | |
| Runway / Pika | Doubling down on video as their only product | Niche specialists with no choice but to make the unit economics work |
What Happens to Sora Users
If you were a Sora user, you have several paths forward, none of them seamless:
- Migrate to ChatGPT Images V2: Your existing credits convert at 5:1. If most of your work was actually still imagery (storyboards, B-roll keyframes), this is fine.
- Move to a video-first specialist: Runway Gen-4 and Pika 2.0 are now the leading commercial options. They're more expensive than Sora was on paper, but they're priced to actually be sustainable.
- Wait for Veo 3 public access: Google's Veo 3 is still research-preview only as of this writing, but a public API is widely expected in Q3 2026. Google has the TPU capacity to underwrite it.
- Self-host an open model: Mochi, CogVideoX, and a handful of community-maintained Llama-Vid forks let you run video generation on your own hardware. Quality is a generation behind, but you're not at the mercy of any provider's compute economics.
The Bigger Lesson for AI Buyers
If you're building a business on top of AI APIs, the Sora shutdown is a reminder that provider product roadmaps are constrained by GPU economics, not by what users want. A model that you depend on can be killed not because it stopped working, but because someone in a finance meeting noticed it was cheaper to redirect the chips.
The defensive play is the same as it has always been:
- Don't build deep dependencies on a single proprietary endpoint. Abstract behind your own thin layer so you can swap providers.
- Track unit economics, not just user love. A product losing the provider money on every call is a product on borrowed time.
- Have an open-source fallback ready, even if it's a generation behind. It's your insurance policy against shutdowns.
- Watch the strategic priorities, not just the marketing. Sora was OpenAI's proudest demo for almost two years - and it was still expendable.
Frequently Asked Questions
When exactly is Sora being shut down?
The API stopped accepting new signups on April 28, 2026. Existing API customers retain access (with reduced rate limits) until July 31, 2026. After that, the endpoint returns 410 Gone.
Why now, after OpenAI invested so much in Sora?
Compute is the binding constraint at every frontier lab. ChatGPT Images V2 turned out to monetize each GPU-second roughly 5–10x better than Sora ever did, and "more chips for the better-monetizing product" is a near-irresistible argument inside any cloud-cost-conscious AI provider.
Is video generation dead?
No - but it's now a niche/specialist market rather than a frontier-lab race. Runway, Pika, and a few open-source projects will continue to push it forward. Expect prices to rise to actually-sustainable levels.
What about Veo 3 from Google?
Google's Veo 3 is technically the strongest text-to-video model still standing. Google's TPU advantage gives them better unit economics on video than OpenAI had on Nvidia hardware. Expect a public API in Q3 2026 if Google decides the consumer/developer market is worth the effort.
Will my Sora credits transfer to images?
Yes - OpenAI is converting Sora credits to ChatGPT Images V2 credits at a 1:5 ratio. One Sora-minute equivalent becomes five ChatGPT Images.
What's next for OpenAI?
The compute freed up from Sora is being redirected to ChatGPT Images V2 capacity, the agentic Codex backend, and (per leaks) GPT-5.5 training. The pattern is clear: bundle into the subscription, take the margin, drop everything that doesn't pay back.
Track current model pricing on our models page - we update it whenever providers move. Use our Token Calculator to estimate the cost of switching workflows from one provider to another.