On February 13, 2026, OpenAI retired GPT-4o—alongside GPT-4.1, GPT-4.1 mini, and o4-mini—from ChatGPT. But the models didn’t vanish entirely. Developers can still call GPT-4o through the API, and the real news is the arrival of GPT-5.6, a family of three enterprise-focused models that redefines how AI fits into corporate Windows environments.
What Actually Changed
The retirement halved the chatter about an abrupt “Omni era” ending. For ChatGPT users, GPT-4o is no longer selectable. For API customers, it lives on—for now—with OpenAI promising advance warning before any future API sunset. That distinction matters for any organization whose tools or scripts speak directly to the API.
Custom GPTs didn’t get stranded. OpenAI automatically moved GPTs that were built on retired models to the nearest GPT-5.3 or GPT-5.4 equivalent. So a custom assistant might wake up with a new brain, but it won’t break overnight.
Meanwhile, Sora—the video-generation tool—has its own separate shutdown. The web and app experiences closed on April 26, 2026, and the API follows on September 24, 2026. Users have a window to export their created content before data is permanently deleted.
The centerpiece, though, is the GPT-5.6 series: Sol, Terra, and Luna. Sol is the flagship, built for complex coding, multi-step agent tasks, and heavy analysis. Terra balances capability and cost, suitable for internal copilots and document workflows. Luna is the speed king, priced for high-volume automation, classification, and extraction.
API pricing per million tokens tells the story: Sol at $5 input and $30 output, Terra at $2.50 and $15, Luna at $1 and $6. A 30-minute minimum cache life and a 90% cached-input discount nudge developers toward smarter prompt architecture.
What It Means for You
For everyday ChatGPT users
If you’re on a free tier, you probably never selected GPT-4o in the first place, so the change may go unnoticed—except that any shared links or conversations now default to a newer model. Paid subscribers (Plus, Pro) can still get Sol’s muscle inside standard ChatGPT, but the conversational warmth many associated with 4o won’t be identical. Tone, response length, and creative style will differ. If you built personal workflows around a specific model’s “feel,” you’ll need to adapt.
For developers and IT pros
Your API integrations didn’t break, but they now have an expiration hanging over them. OpenAI says retired ChatGPT models “continue to be available” through the API—no sunset date yet. Still, this is a signal: port your code to GPT-5.6 sooner rather than later. The tiered lineup gives you real cost control. A chatty internal bot that answers HR questions can run on Luna for pennies, while the code-review agent that inspects pull requests can call Sol when the task demands it.
Programmatic Tool Calling and multi-agent orchestration are the big leap. Instead of generating text and then waiting for another service to parse it, GPT-5.6 can write and execute in-memory programs that co-ordinate tools—search a knowledge base, query a database, draft a message, and file a ticket, all within one request. That turns AI from a smart assistant into a supervised worker. It also means you must design permissions carefully. A model that can call APIs can also call the wrong ones.
For enterprise decision-makers
The realignment is clear: OpenAI is building for managed work, not casual chat. The three-tier structure forces you to think about workload placement. Not every employee needs access to Sol; granting it by default raises cost and risk without proportionate value. But denying Sol to engineers who need to untangle a 10,000-line codebase is penny-wise and pound-foolish.
Security is paramount. All three GPT-5.6 models sit at “High” capability under OpenAI’s Preparedness Framework for cybersecurity and bio risk. Sol and Terra can identify vulnerabilities and stitch together exploits in lab settings, though not yet autonomously against hardened targets. Before turning them loose on internal systems, you’ll need approval gates, human-review steps, and tight integration with your identity platform—something Windows shops can do well, given Entra ID, Intune, and the Microsoft security stack.
How We Got Here
GPT-4o arrived with a multimodal splash, handling text, images, and voice in near-real time. It was the model that made many users feel AI had a personality. By late 2025, however, OpenAI was already signaling that the future belonged to models built for reasoning, coding, and orchestration, not just conversation. In January 2026, the company announced the February 13 retirement date, acknowledging that it had previously restored GPT-4o for Plus and Pro users who valued its creative ideation style.
The industry responded. Google pushed Gemini Omni Flash, a video-centric, consumer-friendly model that slots into YouTube Shorts and the Gemini app, reclaiming the “omni” brand for creative media. Meanwhile, Chinese competitors like ByteDance and Kwai Show ramped up video generation, and Anthropic doubled down on document reasoning. The message: the market is segmenting into enterprise utility and consumer creativity, and OpenAI chose the former.
What to Do Now
Ignore the headlines that declare an abrupt end with no migration paths. Instead, audit your actual usage. Find every place—scripts, Power Automate flows, internal tools, Custom GPTs—that calls a specific model. Build a small evaluation set of real but sanitized tasks. Run them against Sol, Terra, and Luna, measuring not only accuracy but also latency, cost, and tone.
If you use the API, check your logs. GPT-4o calls will keep working, but you should set a calendar reminder to revisit this when OpenAI announces a formal API deprecation. Meanwhile, experiment with prompt caching. If your app sends the same system instructions over and over, a 90% discount on cached tokens can cut bills sharply.
For Sora users, the deadline is harder. Export your library before the API sunsets on September 24, 2026. There is no replacement inside the ChatGPT ecosystem; you’ll need to look at Runway, Pika, or Google’s offerings if video generation is essential.
IT leaders should update their acceptable-use policies. A model that can autonomously chain tools is a force multiplier—and a governance headache. Document which roles get Sol, which stick with Terra, and which workflows are safe to automate fully. Treat the transition as a production change: involve security, compliance, and change advisory boards.
Outlook
The retirement of GPT-4o from ChatGPT is less about one model and more about the end of the general-purpose AI monolith. GPT-5.6 is a portfolio play, and it will be joined by others—including GPT-6, rumored to bring a 2-million-token context window and autonomous agents. The organizations that thrive won’t be the ones that rushed to adopt the newest model. They’ll be the ones that built systems portable enough to swap models when the business case shifts.