OpenAI first flagged the retirement in a ChatGPT release note dated 28 May 2026, giving o3 a ninety-day sunset period that expired today. The change affects ChatGPT only: o3-pro remains listed for Pro, Team, Enterprise and Edu subscribers, according to independent reporting, though OpenAI’s own retirement notice named only the standard o3 model. Two o3 snapshots used through the API — o3-2025-04-16 and o3-pro-2025-06-10 — are on a separate, later clock, with removal scheduled for 11 December 2026 after developers were notified on 11 June 2026. GPT-4.5 was retired from ChatGPT in the same announcement, on a shorter thirty-day sunset that closed in late June 2026.
Why it matters
Anyone who built a ChatGPT habit or workflow around o3’s visible chain-of-thought reasoning — rather than GPT-5.6’s default behaviour — loses that option in the consumer app today, with the paid o3-pro tier as the only route back to broadly similar reasoning. Developers calling o3 through the API have a longer runway, into December, but should treat the ChatGPT retirement as an early warning that the same snapshots are also on notice. The gap between the two deadlines is itself informative: it shows OpenAI is now willing to retire a model from its flagship consumer product months before pulling it from paying API customers.
What changed today
According to OpenAI’s Model Release Notes and ChatGPT Release Notes pages, the retirement note reads: “OpenAI o3 will be retired from ChatGPT on August 26, 2026 following a 90-day sunset period.” OpenAI frames this as part of a routine effort to “retire older models with limited usage in ChatGPT” so it can “better serve” newer, more capable models. That is a vendor characterisation of demand, not an audited usage figure, and OpenAI has not published subscriber-level numbers on how many people were still selecting o3 in the model picker before today.
The practical effect: o3 disappears from the ChatGPT model selector for paid users who previously reached it through the “Show additional models” toggle. Chats that were running on o3 do not simply vanish — OpenAI’s release note states the change applies “to ChatGPT only,” and reporting on user forums describes existing threads rolling over to a GPT-5.6 variant once the old model is no longer selectable, consistent with how OpenAI handled the GPT-4o and GPT-4.1 retirements from ChatGPT earlier in the year. 18BYTE could not independently verify the mechanics of that rollover against a primary OpenAI support document, so it should be read as the reported behaviour rather than a confirmed technical guarantee.
What remains available
o3-pro, the higher-compute variant OpenAI launched in June 2025 “for challenging questions where reliability matters more than speed,” is not named in the May 2026 retirement notice. Several outlets covering today’s cutover report that o3-pro continues to appear in the ChatGPT model picker for Pro, Team, Enterprise and Edu subscribers. 18BYTE was not able to confirm this directly against an updated OpenAI page dated on or after 26 August 2026, so the claim is attributed to that reporting rather than treated as OpenAI’s own confirmed statement; readers on those tiers should check the model picker directly rather than assume continuity.
On the developer side, the API is unaffected by today’s change. OpenAI’s own deprecations page lists two o3-family snapshots — o3-2025-04-16 and o3-pro-2025-06-10 — as scheduled for removal on 11 December 2026, with GPT-5.6 Sol named as the recommended replacement for both. That page also shows a separate, earlier removal for o3-mini-2025-01-31, due 23 October 2026, and for o3-deep-research-2025-06-26, due 23 July 2026 — evidence that OpenAI is working through the o3 family’s various snapshots on staggered timetables rather than retiring the whole lineage at once.
Context: a busier 2026 deprecation calendar
Today’s retirement is not an isolated event. OpenAI’s own release notes show a pattern building through the year: GPT-4.5 left ChatGPT on a thirty-day sunset that closed in late June 2026, in the same notice that set o3’s ninety-day clock; GPT-4o and several 4.1 variants were pulled back from ChatGPT’s default lineup earlier in 2026 before partially returning; and the Assistants API — a separate product from the ChatGPT app — was reported to be shutting down around the same week as today’s o3 cutover. Set against that backdrop, and against a market where rivals are competing hard on inference cost (see 18BYTE’s look at the industry’s cost-per-answer race), the o3 retirement reads as routine portfolio consolidation rather than a one-off decision. OpenAI has not published a consolidated “2026 deprecation calendar,” so this pattern is assembled from separate release notes and deprecation pages rather than a single vendor statement.
The naming scheme is also a signal of where OpenAI wants ChatGPT traffic to land. The recommended replacement for o3 in OpenAI’s own developer documentation is GPT-5.6 Sol — one of three GPT-5.6 variants 18BYTE has covered, alongside GPT-5.6 Luna and GPT-5.6 Terra — rather than a direct o-series successor. That is consistent with OpenAI folding its reasoning-model lineage into the general GPT-5.x naming track inside the ChatGPT app, rather than maintaining o-series branding as a parallel product line.
User and developer reaction
Coverage of today’s cutover, largely from independent trackers and forum monitoring rather than OpenAI’s own channels, describes a vocal minority of o3 users who valued the model specifically for its visible reasoning trace on difficult problems — proofs, subtle code bugs, complex refactors — and who say GPT-5.6’s default behaviour does not reproduce that experience. The same sources report bug complaints in the run-up to retirement, including sessions where long-running o3 chat histories reportedly failed to reload correctly, with affected users advised informally to copy important answers out of ChatGPT before refreshing the page. OpenAI has not issued a public statement addressing these reports, and 18BYTE has not been able to verify them against OpenAI’s own support channels, so they are presented here as user-reported and unconfirmed rather than established fact.
On the enterprise side, commentary framed the retirement as a test of how well organisations that embedded o3 into approval workflows, evaluation pipelines or internal tools have built in model-deprecation handling — a governance question that predates this specific retirement but is sharpened by OpenAI’s shortening cadence between “notice” and “removal” in ChatGPT compared with the API.
What this means for buyers
Consumer and Plus/Pro ChatGPT subscribers who relied on o3 should check the model picker today: if o3-pro is still listed on your plan, it is the closest available substitute, though OpenAI’s own materials describe it as slower per response because of the additional compute it uses. Anyone without access to o3-pro is now defaulted to GPT-5.6 variants for reasoning tasks. Enterprises and developers using o3 through the API have materially more time — removal is not scheduled until 11 December 2026 — but should treat today’s ChatGPT retirement as the first of two deadlines, not a one-off event, and begin testing GPT-5.6 Sol as the vendor-recommended replacement well before the API cutoff. Procurement teams evaluating OpenAI against competitors should also weigh that OpenAI is now retiring consumer-facing models on a ninety-day cycle from announcement, materially faster than the lead times some enterprise buyers may assume.
What would change our reading
A dated, on-the-record OpenAI statement confirming o3-pro’s continued availability past 26 August 2026 — rather than reporting attributed to secondary sources — would let 18BYTE upgrade that claim from Tier 2/3 sourcing to Tier 1. Direct confirmation from OpenAI of what happens to in-progress or archived o3 conversations, and of the reported reload bug, would likewise firm up the “what changed” section. Publication of usage data behind OpenAI’s “limited usage” justification, or a consolidated OpenAI deprecation calendar covering all 2026 retirements, would let this piece move from a reconstructed pattern to a sourced one. If OpenAI updates either release-notes page with a different GPT-4.5 retirement date (the two pages currently show 26 and 27 June 2026 respectively), that discrepancy should be revisited.