The AI Tools That Died in 2026 (And What to Move To)

TL;DR: In 2026, OpenAI discontinued the Sora app (April 26) and will remove the Sora API (September 24); Google shut down three Imagen 4 endpoints (August 17); OpenAI is discontinuing the ChatGPT Atlas browser (August 9) and retiring the DALL-E GPT inside ChatGPT (August 30), after already shutting down the dall-e-2/dall-e-3 API models (May 12). Every date below comes from the vendor's own page, read directly.

Most "AI tools that died" roundups blur four completely different events into one scary word: shutdown. An endpoint shutdown kills one API model ID while the product built on it keeps running under a different model. A product discontinued means the whole application is gone. A feature retired means one named tool inside a living product disappears while the product itself continues. A rename or an acquisition means the thing you used is still operating, just under new branding or new ownership — which is not the same story at all, and lumping it in with a real shutdown either panics people who have nothing to fix or leaves people who genuinely need to migrate thinking they have more time than they do.

Here is what actually happened in 2026, sorted into those categories, with the vendor's own words and the date I read them.

Did Sora shut down in 2026?

Yes, in two separate stages, and this is the one most likely to be reported wrong because the two dates get merged into one. OpenAI's own help center states it plainly: "The Sora web and app experiences were discontinued on April 26, 2026. The Sora API will be discontinued on September 24, 2026" (What to know about the Sora discontinuation, accessed September 3, 2026).

If you used the consumer app to generate clips, that deadline is long past — your task now is exporting anything you still want, at sora.chatgpt.com/sunset. If you called the Videos API directly, sora-2 and sora-2-pro and their dated snapshots stop working on September 24. OpenAI's own deprecations page lists no recommended replacement for any of the six Sora entries — the replacement column is simply blank (API deprecations, accessed September 3, 2026). We already published the full syntax-conversion guide for this one: Sora 2 Shuts Down September 24: Move Your Prompts Now walks through converting Sora prompts to Veo 3.1, Kling 3.0 and Runway Gen-4.5. Link there for the migration mechanics rather than re-deriving them here.

Did Google retire Imagen?

Partly, and it is a textbook endpoint shutdown rather than a product death. Google's own Gemini API deprecation table lists imagen-4.0-generate-001, imagen-4.0-ultra-generate-001 and imagen-4.0-fast-generate-001 with a shutdown date of August 17, 2026 and gemini-3.1-flash-image as the listed replacement (Gemini deprecations, accessed September 3, 2026).

That date is an API retirement, not Google switching off image generation. Consumer surfaces such as the Gemini app had already moved onto the Nano Banana models before this date, and images you already generated with Imagen are unaffected. We covered the full migration — the model mapping, the prompt-style conversion, and the pricing change — in Imagen Is Retired: Migrating Your Prompts to Nano Banana. If Imagen was your image model, that post is the deeper reference; this one is the scorecard entry.

What happened to ChatGPT Atlas?

This is the cleanest product discontinued case in the list, and unlike Sora and Imagen, it is not something we have covered elsewhere yet. OpenAI's own help center article states: "We’re deprecating Atlas and moving browser-based agentic capabilities into ChatGPT and Codex. ... Atlas is scheduled to stop working on August 9, 2026" (Evolving Atlas into ChatGPT for browser-based agentic work, accessed September 3, 2026).

Atlas was OpenAI's own Chromium-based AI browser. The whole application goes away, not one model ID inside it, which is why this is a different category from Sora and Imagen. OpenAI's article is explicit about what does and does not survive the cutoff:

• Bookmarks will not transfer automatically. Export them from Atlas as an HTML file before August 9, then import that file into another browser.
• Open tabs and browsing history will not transfer automatically. Bookmark or copy the URLs you need before the date passes.
• ChatGPT conversation history is separate from Atlas browser data and stays accessible through ChatGPT regardless of what happens to the browser.
• Cookies and signed-in sessions are not portable between browsers at all. Treat any exported session file as sensitive, and never share it.

If your team standardized on Atlas for agentic browsing tasks, OpenAI's own recommended path is the ChatGPT desktop app for multi-tab, download-capable browsing, or the ChatGPT Chrome extension for lighter in-browser help. Neither is a drop-in replacement for a dedicated browser, so budget time to rebuild whatever workflow depended on Atlas specifically, rather than assuming a setting will carry over.

Is DALL-E completely gone?

This is the case that best shows why the four categories in the intro matter, because DALL-E's 2026 story is actually three separate events on three separate dates, and none of them mean "OpenAI stopped generating images."

May 12, 2026 — an endpoint shutdown. OpenAI's deprecations page shows both dall-e-2 and dall-e-3 retired from the API on that date, with gpt-image-2, gpt-image-1 or gpt-image-1-mini listed as the recommended replacements (API deprecations, accessed September 3, 2026). Anything calling those model IDs directly needed to switch by then.

August 30, 2026 — a feature retired inside a living product. OpenAI's own ChatGPT release notes, dated July 31, 2026, read: "On August 30, 2026, we’re retiring the official DALL·E GPT in ChatGPT. We encourage you to download any images you want to keep before then. To continue creating or editing images, use ChatGPT Images. User-created GPTs with image generation enabled are not affected" (ChatGPT release notes, accessed September 3, 2026). Note what that last sentence rules out: this is the one specific, OpenAI-branded GPT going away, not a rule against image generation inside custom GPTs generally.

No date at all — image generation itself. ChatGPT Images, running on the gpt-image family, is the replacement OpenAI names in its own notes above. Nothing in either source suggests OpenAI is stepping back from image generation; the DALL-E name is what is disappearing, model by model and surface by surface, while the capability moves to models with a different name attached.

So "is DALL-E dead" has three honest, different answers depending on which DALL-E you mean. If you have images sitting inside chats with the official DALL-E GPT, download them before August 30. If your code calls dall-e-2 or dall-e-3 directly, that call already started failing on May 12.

Is GPT-4o gone too?

Also worth separating from a full product death, because this is a model retired from one surface, and the surface matters. OpenAI's own announcement states: "On February 13, 2026, alongside the previously announced retirement of GPT‑5 (Instant and Thinking), we will retire GPT‑4o, GPT‑4.1, GPT‑4.1 mini, and OpenAI o4-mini from ChatGPT. In the API, there are no changes at this time" (Retiring GPT-4o, GPT-4.1, GPT-4.1 mini, and OpenAI o4-mini in ChatGPT, accessed September 3, 2026, post dated January 29, 2026).

Read that second sentence carefully, because it is the part every "GPT-4o is dead" headline drops: as of that January 29 announcement, OpenAI drew a line specifically around the ChatGPT consumer interface. The same post says usage had shifted so heavily to GPT-5.2 that "only 0.1% of users" were still choosing GPT-4o each day, which is presumably why the ChatGPT-app retirement moved forward on its own timeline rather than the API's. If your prompts and instructions were written for a named model inside the ChatGPT app specifically, February 13 is the date that mattered to you. If you called the model through the API, check the current deprecations page yourself before assuming anything — a January announcement about "at this time" is not a permanent guarantee months later.

If you want to preserve conversations that depended on a model you can no longer select, OpenAI documents a general export path through the Privacy Portal or ChatGPT's own settings (Exporting your ChatGPT history and data, accessed September 3, 2026) — worth running before you lose the ability to see how a specific model actually responded to your prompts.

The 2026 scorecard, at a glance

A copy-paste audit for your own codebase and prompt library

Before you assume you are unaffected by any of the above, check what you actually call rather than what you remember calling. This is the same five-minute audit worth running against any vendor's deprecation list, not just this one:

# Search your codebase, notebooks and cron jobs for pinned model IDs
grep -rniE "sora-2|dall-e-2|dall-e-3|imagen-4\.0-(generate|ultra-generate|fast-generate)-001|gpt-4o\b" .

# Search your saved-prompt library or notes app for the same strings
# (do this manually if your prompts live outside version control —
#  that is exactly the gap a portable prompt library closes)

A single missed call site against a retired model ID typically fails outright at the request layer rather than degrading quietly, so this is worth running now rather than after something breaks in production.

What to actually do about it

Five things, in order of how much time they save you the next time this happens, and it will happen again: gemini-2.5-flash-image already has an October 2026 date attached to it.

1. Get your prompts out of chat history and into something you own. A chat log is not a prompt library. If the only copy of your best prompt lives three scrolls up in a conversation with a model that just disappeared, you have already lost the fast path to recovering it.
2. Pin model names as data, not as buried strings. A grep-able config value survives a deprecation notice. A model name hardcoded in forty places does not.
3. Export before a deadline is announced, not after. Every case above had an announced cutoff with weeks or months of notice. Treat "export available" as a standing habit, not a fire drill.
4. Read the vendor's own deprecation page, not a roundup of it. Every example above required distinguishing an endpoint date from a product date from a feature date, detail that gets lost the moment a summary article compresses five sources into one sentence.
5. Keep your working prompts platform-agnostic where you can. A prompt written as portable instructions, with the model-specific syntax layered on top, survives a migration in minutes. A prompt written entirely in one vendor's dialect does not.

Step one is where a real prompt library earns its keep, because the alternative is reconstructing your best work from memory the week a tool you depend on gets a retirement notice.