OpenAI Retires o3 From ChatGPT Tomorrow, Closing the Book on the o-Series Era

Tomorrow, August 26, 2026, one of the most consequential models in AI history quietly disappears from ChatGPT. OpenAI’s o3 — the reasoning model that shattered the ARC-AGI benchmark in December 2024 and redefined what “thinking” looked like in a consumer product — will be retired from the ChatGPT model picker following a 90-day sunset period. The API is unaffected, for now. But in the app where hundreds of millions of people actually touch these models, the o-series era is officially over.

What’s happening

According to OpenAI’s Model Release Notes, o3 will be retired from ChatGPT on August 26, 2026, following a 90-day sunset window that began in late May. GPT-4.5 was the domino before it, vanishing on June 27, 2026 after a shorter 30-day sunset — a retirement that ended the entire GPT-4 era in ChatGPT. With o3 now going dark, every model in the ChatGPT picker descends from the GPT-5 lineage, and specifically from the GPT-5.6 family released in July 2026.

It’s worth pausing on how fast this happened. o3 was announced in December 2024 as the grand finale of AI’s year — a system that scored 75.7% on the ARC-AGI public set (and 87.5% in its high-compute configuration), a benchmark that GPT-4o could barely crack at 5%. ARC Prize, the organization behind the benchmark, called it a breakthrough score. For a few weeks, o3 was the closest thing the industry had to a public AGI moment. Twenty months later, it’s being sunset from the product that made it famous.

Why o3 mattered

o3 belonged to a generation of models that made “reasoning” a product feature. Instead of answering immediately, the model would think — sometimes for minutes — chaining intermediate steps before committing to an answer. That test-time-compute approach delivered step changes on math, code, science, and logic benchmarks, and it forced every competitor to respond: Google’s Gemini thinking modes, Anthropic’s extended thinking, DeepSeek’s R1, and the entire open-weight reasoning wave that followed.

The strategic significance went further. o3 was OpenAI’s proof that scaling inference-time compute — not just training compute — could buy capability. That insight is now baked into everything: the reasoning-effort sliders in ChatGPT, the “thinking budgets” in competing APIs, and the benchmark culture where ARC-AGI-2 and ARC-AGI-3 scores are reported at multiple compute settings precisely because the industry learned from o3 that the knob matters.

But o3 was also a standalone artifact of an architectural moment: a reasoning model built as its own product line, separate from the “base” GPT models. The GPT-5 generation merged those lineages. Reasoning stopped being a separate model you picked and became a dial on every model. Once that happened, the o-series was living on borrowed time.

What replaces it

OpenAI’s recommended migration path is GPT-5.6 Sol, the flagship of the GPT-5.6 family that launched July 9, 2026 alongside siblings Terra and Luna. Sol powers the higher reasoning tiers in ChatGPT (Plus users get it on Medium and High reasoning), and OpenAI positions it as state-of-the-art across coding, science, and knowledge work. For users who relied on o3’s deep multi-minute reasoning, OpenAI points to GPT-5.6 Sol on Pro with extended reasoning as the spiritual successor.

The family structure tells its own story about how the product has matured: Sol for frontier reasoning, Terra as the balanced workhorse, Luna as the fast, cheap default — plus automatic model switching, a dedicated Work mode, and user-facing reasoning controls that let you decide how hard the model thinks. That last feature is arguably o3’s real legacy: what was once a dramatic, separate “reasoning model” is now a configuration option on everything.

Migration guidance published across OpenAI’s help center and third-party checklists is consistent: teams that built workflows around o3’s behavior should re-validate them on GPT-5.6 Sol before tomorrow, paying attention to reasoning-effort settings, latency profiles, and output formatting differences. API users have more runway — o3’s API availability is unchanged by this retirement — but the direction of travel is unmistakable.

The pattern: ChatGPT as the new retention frontier

Zoom out, and o3’s retirement is the latest data point in an aggressive lifecycle policy that OpenAI has applied all year. GPT-4o, GPT-4.1, and GPT-4.1 mini were retired from ChatGPT on February 13, 2026. GPT-4.5 followed on June 27. o3 on August 26. Each retirement follows the same script: sunset announcement, migration window, removal from the app, API untouched — for now.

For a company that once kept models around for years, the pace is striking, and it reflects two pressures. First, operating economics: every model in the picker is a serving cost, and consolidating traffic onto a smaller, newer set of models simplifies routing, caching, and capacity planning across what is now a globe-spanning inference fleet. Second, product coherence: the GPT-5.6 architecture — one family, three sizes, a reasoning dial — is easier to explain, price, and improve than a museum of historical model names, and it nudges users toward the automatic-switching behavior OpenAI prefers over manual model selection.

There’s a cultural cost, though, and Reddit threads mourning “the last two good models” capture it: power users formed genuine preferences for specific models, and each retirement forces a re-learning cycle. The model picker is becoming less like an archive and more like a conveyor belt.

What to watch

• August 26, 2026: o3 disappears from ChatGPT. Watch for OpenAI to mark the moment — or, more likely, say nothing, consistent with how quietly these sunsets have been executed.
• API timeline: o3 remains available via API, but GPT-5-era deprecations (GPT-5 and o3 API access was shut down March 26, 2026 per OpenAI’s deprecations page for certain tiers) suggest the API clock is also ticking. Builders with long-lived o3 integrations should have a GPT-5.6 migration branch ready.
• The benchmark era o3 defined: ARC-AGI has since moved to versions 2 and 3, where even GPT-5.6 Sol scores remain modest on the hardest puzzle sets. The frontier o3 briefly cracked open has receded as benchmarks hardened — a reminder that “breakthrough” is always measured against a moving target.
• DALL-E GPT interface: also retiring August 30, 2026, per recent reports — the cleanup of legacy surfaces continues.

The takeaway

o3’s retirement is not a failure story — it’s a succession story. The model that made reasoning a headline feature is being replaced by an architecture where reasoning is simply assumed, a dial rather than a destination. That’s how progress looks from inside the product: the revolutionary thing stops being a product and becomes a property of everything.

For the researchers who cheered the December 2024 ARC-AGI moment, tomorrow is a small funeral. For everyone else, it’s Tuesday, and ChatGPT will feel exactly the same — which is precisely the point.