Somewhere on Reddit right now there's a post that reads like a eulogy. Not for a person — for a chatbot. Users mourning GPT-4o described it as more than software: one person wrote it "was part of my routine, my peace, my emotional balance," and insisted on calling it "him" because it never felt like code. That reaction tells you everything about what OpenAI dismantled in 2026, and why the story doesn't end with a quiet model deprecation. It ends with Google walking on stage months later and calmly claiming the word "Omni" for itself.
Here's the full timeline, what actually replaced the Omni lineage, and what it means for anyone still holding out hope for a "GPT-6o."
What actually happened to GPT-4o and the omni lineage?
The dismantling wasn't a single event — it was a slow-motion demolition spread across most of 2026. Here's the dated sequence:
February 13, 2026 — GPT-4o's first cut in ChatGPT. OpenAI confirmed on February 13, 2026, alongside the previously announced retirement of GPT‑5 (Instant and Thinking), it retired GPT‑4o, GPT‑4.1, GPT‑4.1 mini, and OpenAI o4-mini from ChatGPT. At the time, there were no changes to the API.
February 16-19, 2026 — The API door started closing too. Business, Enterprise, and Edu customers got a short grace period, but GPT-5.1 models were retired on March 11, 2026 across both normal chats and GPTs, though they'd continue to be available through the OpenAI API with advance notice ahead of future retirements.
April 3, 2026 — Custom GPTs lost their last thread to 4o. ChatGPT Business, Enterprise, and Edu customers retained access to GPT-4o within Custom GPTs only until April 3, 2026, after which GPT-4o was fully retired across all plans.
April 26, 2026 — Sora's web and app died. The Sora web and app experiences were discontinued on April 26, 2026, with the Sora API following on September 24, 2026. If you had unsaved projects, OpenAI's help page notes you could export content by going to sora.chatgpt.com/sunset and clicking Export, but after that window, everything gets permanently deleted.
June 26, 2026 — Even GPT-4.5 wasn't spared. As of June 26, 2026, GPT-4.5 is no longer available in ChatGPT, including for custom GPTs, though existing conversations that used GPT-4.5 can continue with GPT-5.5.
You can read the official retirement notices directly on OpenAI's Help Center and the original retirement announcement.
Why did OpenAI retire GPT-4o if users loved it so much?
Two very different explanations are floating around, and both are probably true at once.
The official line is usage math. Reporting around the retirement repeatedly cites the same number: only 0.1% of ChatGPT usage was still choosing GPT-4o daily, with OpenAI citing this adoption metric as the primary driver behind the retirement decision.
The messier explanation is legal and emotional. GPT-4o wasn't a neutral tool — it was, for a specific slice of users, an emotional fixture. When OpenAI first tried replacing it with GPT-5 in August 2025, massive backlash forced them to reverse course, which is exactly why the model stuck around as long as it did. But the February 13, 2026 retirement came as the company faced eight lawsuits alleging the model's overly validating responses contributed to mental health crises. OpenAI's own retirement post admits as much in softer language, acknowledging it brought GPT-4o back after hearing feedback from a subset of Plus and Pro users who preferred its conversational style and warmth — before ultimately deciding that warmth wasn't worth the liability.
What replaced GPT-4o, and is it any good?
If you're still on ChatGPT day-to-day, the transition was mostly invisible. After the cutoff date, these models no longer appear in the ChatGPT model selector, and all existing conversations and custom GPTs default to GPT-5.2. OpenAI tried to soften the personality gap too — it introduced customizable personality settings in GPT-5.2 to address user feedback about GPT-4o's conversational tone.
But the real successor arrived months later, and it wasn't a single model at all.
What is GPT-5.6, and why does it come in three flavors?
On July 9, 2026, OpenAI shipped what it's calling its most capable system yet — and abandoned the old naming convention entirely. The names are a deliberate change from the old "mini" and "nano" labels. Instead, you get three durable capability tiers: Sol, Terra, and Luna.
Sol is the new flagship, Terra is a balanced model for everyday work, and Luna is the most cost-efficient option. Pricing breaks down clearly: Luna runs $1/$6, Terra $2.50/$15, and Sol $5/$30 per million input/output tokens. The efficiency story is arguably the bigger deal for most users — OpenAI says Terra runs at competitive performance to GPT‑5.5 while being 2x cheaper, which matters way more for people running high-volume workflows than any single benchmark jump.
On the benchmark side, GPT-5.6 Sol set a new high of 53.6 on Agents' Last Exam, eclipsing Claude Fable 5 by 13.1 points, though it's not a clean sweep — Fable 5 got 80% on SWE-Bench Pro compared to GPT-5.6 Sol's 64.6%. You can dig into the full release notes on OpenAI's GPT-5.6 announcement page or the more technical preview post for Sol.
If you're a developer, GPT-5.6 is already live in GitHub Copilot, where it comes in three variants so you can match the model to the job, whether that's reasoning over a large codebase, everyday agentic coding, or fast, cost-efficient assistance.
Why did Google claim the "Omni" name for itself?
Here's the twist that makes this whole story worth telling. Three months after OpenAI buried GPT-4o, Google stood on stage at I/O 2026 and unveiled a product family with the exact same word in its name.
At Google I/O 2026, Google released two new models, Gemini Omni and Gemini 3.5, with Gemini Omni able to create anything from any input starting with video, and Gemini 3.5 Flash the first in its latest family combining frontier intelligence with action. It wasn't buried in a footnote either — the core structural development of the conference centered on the debut of Gemini Omni Flash, the first release in a new "Omni" family of world models.
Rollout was aggressive and, notably, partly free. Gemini Omni Flash rolled out immediately to all Google AI Plus, Pro and Ultra subscribers globally through the Gemini app and Google Flow, and was also available at no cost in YouTube Shorts Remix and the YouTube Create app for users 18+. You can browse the full rundown on Google's official I/O 2026 announcement collection.
What does Gemini Omni Flash actually do differently?
Under the hood, Google essentially collapsed several separate products into one model. According to independent analysis, a single unified stack means Omni collapses what used to be Veo (video) + Imagen (image) + separate audio systems into one model, which should reduce cross-modality artifacts. Every output also carries provenance tagging — Google says Omni videos include SynthID digital watermarking, while recent tests show prompts can still push the model toward highly recognizable IP-style characters.
Worth noting: this isn't a finished product yet in the way GPT-5.6 is. Google hasn't published numeric benchmarks alongside the launch, so independent evaluation is still pending — the most consequential follow-up will come when the API opens up.
So is there actually a "GPT-6o," or was that always a myth?
This is the part that trips people up, and the honest answer is: no, not yet — and possibly never with that exact name. What a lot of people expected to ship as "GPT-6" actually landed months earlier as GPT-5.5. What was widely expected to ship as GPT-6 actually shipped on April 23, 2026 as GPT-5.5, codenamed "Spud" during development, with the memory and personalization features Sam Altman teased for "GPT-6" landing in 5.5 instead.
That leaves real GPT-6 as something OpenAI is deliberately holding back. One technical analysis frames it well: by reserving the "6" label, OpenAI is betting that GPT-6 will be a qualitatively different product — one where long-term memory is core infrastructure, not a plugin, and where agentic execution is reliable enough to run unsupervised on multi-day workflows. Altman's own public comments back that up directly — he envisions ChatGPT having "infinite, perfect memory," remembering every document, email, and conversation if the user opts in, believing this deep personalization will become OpenAI's true competitive moat beyond just model intelligence.
So the "O" is retired. GPT-6, whenever it lands, is being saved for something OpenAI considers a genuinely different category of product — not another point release with a friendlier voice.
What should you actually do if you relied on any of these tools?
If your workflow depended on GPT-4o, Custom GPTs built around it, or Sora, here's the practical path forward:
- Check the OpenAI ChatGPT release notes regularly — deprecation dates keep moving and OpenAI updates this page directly.
- Rebuild Custom GPTs on GPT-5.2 or later rather than assuming legacy behavior will persist; the personality settings in newer models are OpenAI's attempt to recreate 4o's tone.
- Export anything you still need from Sora before any final export window closes — once it's gone, OpenAI's discontinuation FAQ confirms the data is permanently deleted.
- If you're a developer choosing a new model, weigh GPT-5.6 Luna for cost-sensitive routine tasks against Sol for genuinely hard, long-running agentic work — Terra is the safe middle ground most teams should start with.
- Keep an eye on Gemini Omni's API rollout if multimodal video generation matters to your work — the consumer version is live now, but developer access is still expanding.
The bigger lesson here has nothing to do with model names. It's that the tools people get emotionally attached to are rarely the ones companies keep around the longest — profitability, legal exposure, and usage stats will always win that argument eventually. The question worth sitting with isn't which model has the best benchmark score this month. It's what you're building that can survive the next retirement notice.