OpenAI o3 Leaves ChatGPT on August 26: The Last Standalone Reasoning Flagship Says Goodbye

Five days. On August 26, 2026, Ope​nAI o3 — the reasoning model Ope​nAI released on April 16, 2025 alongside Ope​nAI o4-mini as the company's "smartest models released to date" — disappears from the model picker in ChatGPT's web and mobile apps. It is not the only thing happening to the model, and it is not the loudest. The retirement date is the one with the hard calendar: Ope​nAI announced it in the model release notes on May 28, giving o3 a 90-day sunset period while folding the announcement into the same roster consolidation that already removed Ope​nAI o4-mini from ChatGPT in February and GPT-4.5 in June. But the reason o3's remaining fans are talking about it this week is the other story: a pattern of o3 responses that seem to vanish after a page reload, filed repeatedly with Ope​nAI on its official community forums, and still unexplained as of this writing.

If you are not a ChatGPT power user who manually selects o3, none of this touches you — the model has not been a default option for a while, and Ope​nAI says only a fraction of a percent of daily users still pick it. If you are one of the people who does, or a developer who calls o3 through the API, the next two months matter more than the next five days, and the details below are the ones worth having.

What exactly gets retired, and when

The May 28 announcement, carried in Ope​nAI's model release notes, said the company was "retiring older models with limited usage in ChatGPT to better support newer, most capable models." For Ope​nAI o3 specifically that means the ChatGPT interface, not the API:

• ChatGPT web and mobile — o3 leaves the model picker on August 26, 2026, the end of the 90-day sunset. Until then, paid subscribers can still select it manually; it just stopped being a default long ago.

• The API — the retirement announcement explicitly left the API untouched. A separate developer notice on June 11, 2026 set API removal dates for two o3 snapshots, o3-2025-04-16 and o3-pro-2025-06-10, on December 11, 2026. So developers get a longer runway than ChatGPT users, but not an indefinite one.

• o3-pro — the bigger-compute version remains in ChatGPT for Pro, Team, Enterprise and Edu subscribers. The o3 line is not fully gone; the consumer-grade o3 is what is closing.

The same consolidation wave explains the pattern. Ope​nAI o4-mini, o3's smaller launch-mate, was retired from ChatGPT on February 13, 2026 after a January 29 notice. GPT-4.5 followed on June 27 after a 30-day sunset. o3's 90-day window was the longest of the three, which is consistent with how much harder it is to replace a reasoning model than a writing model — the thing o3 did well is the thing that is hardest to transplant.

Why o3 still has fans in the first place

o3's reputation was earned in the year between its April 2025 launch and the first retirement notices. It was the flagship of the o-series: trained to think before responding, and on the benchmarks Ope​nAI published at launch it posted numbers that became the reference points for "hard reasoning" — 96.7% on AIME 2024, 87.7% on GPQA Diamond, a 2727 Codeforces Elo, and 87.5% on ARC-AGI at high compute, all figures reported by Ope​nAI and, in o3's case, unusually durable against independent retesting in 2025. On SWE-bench Verified, Ope​nAI claimed 69.1% without custom scaffolding. The model could also "think with images," integrating screenshots and diagrams into its chain of thought before it answered.

The fans who stuck with it describe the experience in a way that is hard to translate into a benchmark: o3 explained its reasoning out loud, and for careful, high-stakes work — a tricky proof, a subtle code bug, a complicated refactor — the visible reasoning was the product. Those same users are the ones filing the bug reports this month.

The disappearing-responses bug, and what users report

Across several long-running threads on Ope​nAI's community forums, paid users describe a consistent failure mode: they get a good o3 answer, reload the page or switch away and back, and the response is gone — replaced by nothing, by "There was an error generating a response," or by an earlier state of the conversation. In the more detailed reports, the chat window goes blank after a post, the selected model flips to something else (one user saw o3-mini-high silently become a non-reasoning model), and history only comes back after clicking around. Some threads describe the last several responses disappearing together; others describe a "global reset" where the thread rolls back to a much earlier message.

These are user reports, not an Ope​nAI acknowledgment. Ope​nAI has not confirmed a root cause or published a fix for this class of bug as of this writing, and the threads mix o3 and other models. The pattern that does recur across the reports is the one that matches the retirement timeline: long chat histories, session reloads, and the o3/o3-mini-high endpoints specifically. The practical advice users have settled on — copy important answers out of ChatGPT before reloading, or use the mobile app where the conversation often still renders — is exactly the advice you would give yourself if you assumed the web UI could not be trusted with your last o3 response. That is a grim place for a paid product to be five days before a retirement, and it is one more reason the migration question is not just about "what comes next."

What actually survives o3's retirement

For API users, the honest answer is: the model, for now. o3-2025-04-16 is still callable, and the developer notice gives it until December 11, 2026. That is not "forever," and the June 11 notice is a warning that the API roster is being consolidated toward the same place the ChatGPT roster went — which is why teams building new integrations against o3 today should treat it as a temporary dependency, not a platform.

Inside ChatGPT, the successor path Ope​nAI is pushing is the GPT-5 family. GPT-5.5 (Instant, Thinking, and Pro tiers) has been the default since May, and GPT-5.6 is reportedly in development. For the o3 faithful, the closest thing to the old experience inside Ope​nAI's own product is o3-pro, which keeps the deliberate reasoning behavior for Pro and higher subscribers. The retirement of the cheap o3 tier is effectively an upsell funnel for anyone who wants to keep reasoning in ChatGPT without leaving Ope​nAI.

If you are an API user, the clock looks different

Developers who built on o3 via the API have a real but bounded runway, and the decision that matters is what to standardize on next. o3 was never the budget option — even after Ope​nAI cut its price roughly 80% in June 2025, from $10 to $2 per million input tokens and from $40 to $8 per million output, reasoning tokens are billed as output and the effective cost of a long-thinking answer adds up fast. The models that replaced it at the top of the independent leaderboards — Claude Opus 5, DeepSeek V4 Pro, Gemini 3.1 Pro, GLM 5.2 — are all current, all tested, and none of them is going anywhere.

That is where OrcaRouter comes in, and it is worth being precise about what we do and do not offer. Ope​nAI o3 itself is not on OrcaRouter's key — it remains available through Ope​nAI's own API and several third-party platforms. But every serious successor to it is. One OrcaRouter API key covers 200+ models with the provider's list price passed through at 0% markup, so a vendor price cut is live on our side the same day it is live at the vendor, and automatic failover means a production path can point at two or three o3 replacements and never go down while the roster settles. If you are spending this month deciding where o3's workload goes, that is the argument for testing the replacements through a single endpoint rather than building five separate integrations you will have to maintain.

What the next five days are for

If you are still selecting o3 in ChatGPT, the practical list is short. Export anything in long o3 threads you cannot afford to lose, given the reload bug — copy the important answers out now rather than trusting the conversation to survive. Decide whether o3-pro is worth the subscription tier it sits behind. And if you are an API user, treat the December 11 cutoff as the real deadline: the model works until then, but every week of new integration built against it is a week you will spend migrating. The retirement itself is not a surprise — it was announced in May, it has been on the calendar for three months, and the only genuinely new thing this week is the bug reports making the case that even the model's remaining fans should stop relying on the ChatGPT web app to hold their o3 work. The countdown is short, and unlike the bug, the date is not going to change.