A lot has been happening in the world of AI since ChatGPT hit the scene back in late 2022. Every year, something new and improved came out, offering bigger and better results. Visuals were more refined, text and research were more reliable, voices were more lifelike, and prompt adherence quickly got better. We saw AI reach into more and more industries, promising to improve and speed up workflows.
But now it’s 2026, and we’re seeing a lot of companies shutting down major projects. Some people see it as a sign of the AI bubble bursting. If you don’t look any further than the fact that a “thing shuts down,” it’s understandable that you’ll see it that way.
As usual, there’s more to it than that.
Models get retired all the time in favor of new ones, and services frequently shut down when alternatives are presented. You’ll see all of that happening in the charts below.
In the case of Sora, things are a bit different.
Goodnight Sora

Sora has been a hit or miss with me in the past. When the first version was released, it was fairly good at adhering to prompts for images, especially compared to many off-the-shelf models. Users quickly started generating some highly imaginative images using Sora 1.
However, AI video was where the real power was. Or at least, it was supposed to be.
I may write up a blog post on just how bad the Sora 1 videos were, especially in their attempts at realistic human movements and respect for solid objects like walls. About 30% of the time, I was able to get an output that worked for me and didn’t need additional video editing. Still, image creation was very useful, and I was able to incorporate it into my workflow.
Sora 2, however, was not like that. It was a big improvement in many ways, now including longer clips, mostly usable extended scenes, multi-shot angles, and lip-synced generated voices. It still had a lot of flaws and weird adherence issues, plus a safety filter that made action scenes feel like you were watching a video game. The potential was there, and if they had provided the ability to buy individual HD downloads without the watermark, that would have been excellent.
But they didn’t. Instead, they’re shutting it down entirely. Here’s why:
OpenAI uses a ton of computing power to run Sora. And that computing power is not free. Some estimates put the cost of running Sora at $1 million per day, others have it much higher. For a business to spend that kind of money, it would need to be generating substantially higher revenue from the product.
They aren’t. It’s actually nowhere near enough to cover the cost.
So while they spent a truckload of money to generate goofy meme videos, the return on investment was a fraction of the cost.
So, that’s kind of a problem. You can’t really run a business that way.
While the Sora system is not being “deleted,” it is being shuttered and moved internally. The computing power is being tasked to their other models, including Codex, the coding model used by developers.
More AI shutdowns in 2026
Sora isn’t the only thing going away. There are a lot of AI services, products, and APIs taking a dirt nap this year. However, most are being retired in favor of new models or services. When Apple or Microsoft releases a new operating system, they retire the old one. This isn’t new in the software industry, AI, or any other industry.
Below is a large list of AI tools that have already shut down or will shut down this year, along with any planned path forward by their respective companies.
Big product / API shutdowns
| Company | Service / API | Shutdown date | Notes |
|---|---|---|---|
| OpenAI | Sora web and app | April 26, 2026 | OpenAI Help says Sora web/app experiences will be discontinued. (OpenAI Help Center) |
| OpenAI | Sora API / Videos API / Sora 2 models | September 24, 2026 | Includes Videos API, sora-2, sora-2-pro, and dated Sora 2 snapshots. (OpenAI Developers) |
| OpenAI | Sora 1 legacy experience | March 13, 2026 in the U.S. | Sora 1 became unavailable in the U.S.; Sora 2 became the default. (OpenAI Help Center) |
| OpenAI | Assistants API | August 26, 2026 | Replacement: Responses API and Conversations API. (OpenAI Developers) |
| OpenAI | Realtime API Beta | May 7, 2026 | Replacement: GA Realtime API. (OpenAI Developers) |
| Google / Firebase AI Logic | Imagen models via Firebase AI Logic | June 24, 2026 | Firebase says all Imagen models are deprecated and will shut down; replacement is Gemini Image / “Nano Banana” models. (Firebase) |
| Vertex AI Generative AI module in Vertex AI SDK | June 24, 2026 | Google is steering users to the Google Gen AI SDK. This is an SDK module shutdown, not a model shutdown. (GitHub) |
OpenAI API model shutdowns
| Shutdown date | Models / systems |
|---|---|
| February 12, 2026 | codex-mini-latest; replacement gpt-5-codex-mini. (OpenAI Developers) |
| February 17, 2026 | chatgpt-4o-latest; replacement gpt-5.1-chat-latest. (OpenAI Developers) |
| March 26, 2026 | gpt-4-0314, gpt-4-1106-preview, gpt-4-0125-preview, plus aliases such as gpt-4-turbo-preview. (OpenAI Developers) |
| May 7, 2026 | Realtime/audio preview models: gpt-4o-realtime-preview, gpt-4o-mini-realtime-preview, gpt-4o-audio-preview, gpt-4o-mini-audio-preview, and dated realtime snapshots. (OpenAI Developers) |
| May 12, 2026 | dall-e-2, dall-e-3; replacements gpt-image-1 or gpt-image-1-mini. (OpenAI Developers) |
| August 26, 2026 | Assistants API. (OpenAI Developers) |
| September 24, 2026 | Sora 2 / Videos API models: sora-2, sora-2-pro, sora-2-2025-10-06, sora-2-2025-12-08, sora-2-pro-2025-10-06. (OpenAI Developers) |
| September 28, 2026 | gpt-3.5-turbo-instruct, babbage-002, davinci-002, gpt-3.5-turbo-1106. (OpenAI Developers) |
| October 23, 2026 | Fine-tuned versions: ft-gpt-3.5-turbo, ft-gpt-4, ft-gpt-4.1-nano-2025-04-14, ft-babbage-002, ft-davinci-002. (OpenAI Developers) |
Also, a large number of legacy models will be shut down in July.
https://developers.openai.com/api/docs/deprecations#2026-04-22-legacy-gpt-model-snapshots
Anthropic Claude API retirements
| Retirement date | Models |
|---|---|
| January 5, 2026 | claude-3-opus-20240229. (Claude) |
| February 19, 2026 | claude-3-7-sonnet-20250219, claude-3-5-haiku-20241022. (Claude) |
| April 20, 2026 | claude-3-haiku-20240307. (Claude) |
| June 15, 2026 | claude-sonnet-4-20250514, claude-opus-4-20250514. (Claude) |
Google Gemini / Imagen / Veo model shutdowns
Google’s dates differ by platform: Gemini Developer API, Vertex AI, and Firebase AI Logic do not always have identical schedules.
| Platform | Shutdown / retirement date | Models |
|---|---|---|
| Gemini API | March 9, 2026 | gemini-3-pro-preview. (Google AI for Developers) |
| Gemini API | March 31, 2026 | gemini-2.5-flash-lite-preview-09-2025. (Google AI for Developers) |
| Gemini API | June 1, 2026 | gemini-2.0-flash, gemini-2.0-flash-001, gemini-2.0-flash-lite, gemini-2.0-flash-lite-001. (Google AI for Developers) |
| Gemini API | June 17, 2026 | gemini-2.5-pro, gemini-2.5-flash. (Google AI for Developers) |
| Gemini API | July 22, 2026 | gemini-2.5-flash-lite. (Google AI for Developers) |
| Gemini API | October 2, 2026 | gemini-2.5-flash-image. (Google AI for Developers) |
| Vertex AI | June 1, 2026 | gemini-2.0-flash-001, gemini-2.0-flash-lite-001. (Google Cloud Documentation) |
| Vertex AI | June 30, 2026 | Imagen: imagen-4.0-generate-001, imagen-4.0-fast-generate-001, imagen-4.0-ultra-generate-001, imagen-3.0-generate-002, imagen-3.0-generate-001, imagen-3.0-fast-generate-001, imagen-3.0-capability-001; Veo: veo-3.0-generate-001, veo-3.0-fast-generate-001, veo-2.0-generate-001. (Google Cloud Documentation) |
| Vertex AI | October 2, 2026 | gemini-2.5-flash-image. (Google Cloud Documentation) |
| Vertex AI | Not before October 16, 2026 | gemini-2.5-pro, gemini-2.5-flash, gemini-2.5-flash-lite. (Google Cloud Documentation) |
Cohere retirements
| Date | Models / features |
|---|---|
| April 4, 2026 | embed-english-v2.0, embed-english-light-v2.0, embed-multilingual-v2.0, c4ai-aya-expanse-8b, c4ai-aya-vision-8b. (Cohere Documentation) |
Microsoft Foundry / Azure AI Foundry retirements
Microsoft’s Foundry schedule is platform-specific, so these are retirements inside Microsoft Foundry / Azure, not necessarily the provider’s own native API.
| Provider section in Foundry | 2026 retirements I found |
|---|---|
| Azure OpenAI | gpt-5-chat preview versions May 13; gpt-5.1-chat May 13; gpt-5.2-chat May 13 / June 9 depending version; gpt-5.3-chat June 3; gpt-image-1 May 15; gpt-4o-mini-transcribe, gpt-4o-mini-tts, gpt-4o-transcribe June 1; tts, tts-hd, whisper June 18; o1 July 15; o3-mini August 2; gpt-4o 2024-11-20 October 1; o3 October 16; codex-mini November 15; several GPT/audio/image variants in December. (Microsoft Learn) |
| xAI on Foundry | grok-3, grok-3-mini, grok-4-fast-non-reasoning, grok-4-fast-reasoning retire May 1, 2026. (Microsoft Learn) |
| Cohere on Foundry | Cohere-command-r-08-2024, Cohere-command-r-plus-08-2024 retire May 12; Cohere-rerank-v3.5 retires May 14. (Microsoft Learn) |
| Deci AI on Foundry | deci-decidiffusion-v1-0 retires July 31. (Microsoft Learn) |
| Meta on Foundry | Llama-3.2-11B-Vision-Instruct, Llama-3.2-90B-Vision-Instruct, Meta-Llama-3.1-405B-Instruct, Meta-Llama-3.1-8B, Meta-Llama-3.1-8B-Instruct retire June 13. (Microsoft Learn) |
| Microsoft models on Foundry | financial-reports-analysis, financial-reports-analysis-v2, supply-chain-trade-regulations, supply-chain-trade-regulations-v2 retire July 31. (Microsoft Learn) |
Databricks Mosaic AI Model Serving
| Date | Models |
|---|---|
| March 9, 2026 | Meta Llama 4 Maverick retired for pay-per-token. (Databricks Documentation) |
| March 26, 2026 | Google Gemini 3 Pro retired, with temporary redirection to Gemini 3.1 Pro until June 7. (Databricks Documentation) |
| June 9, 2026 | Meta Llama 4 Maverick retired for provisioned throughput. (Databricks Documentation) |
| Already gone by April 2026 | Anthropic Claude 3.7 Sonnet no longer available on Databricks. (Databricks Documentation) |
Together AI serverless model removals
Together has a long deprecation history. Removals for 2026 include these groups:
- Qwen3 VL / thinking models
- Mixtral 8x7B
- ServiceNow Apriel thinker models
- GLM 4.5/4.7 models
- Mistral Small 24B
- Llama 4 Maverick
- Mxbai Rerank Large V2
- Kimi K2 variants
- Llama 3.1/3.2 models
- FLUX dev models
- Qwen2.5/Qwen3 models
- Salesforce Llama-Rank
- BGE/GTE embedding models
- plus several Together/Refuel models
Dates listed include January 5, February 3, February 6, February 25, March 6, March 31, April 2, April 3, and April 16, 2026.
https://docs.together.ai/docs/deprecations
Perplexity API
| Date | Models |
|---|---|
| March 20, 2026 | google/gemini-2.5-flash removed from Perplexity Agent API. (Perplexity) |
| April 1, 2026 | google/gemini-2.5-pro and google/gemini-3-pro-preview removed from Perplexity Agent API. (Perplexity) |
More AI Closures in 2026
This is not a comprehensive list of closures. It’s only the end of April as I’m writing this. Things can change throughout the year, new models and services can pop up, and old ones can be retired. It’s very likely to happen.
That said, AI isn’t going away. It will evolve as anything else does. Hopefully, one day, OpenAI will bring Sora back for public use, but right now, that seems unlikely.
If they do, Sora 3 will need to be an actual game-changer. Something so strong it smacks down the competition like a seven-fingered Stable Diffusion 1.5 character.
