OpenAI sets o3 and GPT-4.5 ChatGPT sunset dates

AI 요약

• What happened: OpenAI's May 28, 2026 ChatGPT release note paired a GPT-5.5 Instant quality update with model and workspace cleanup.
  • OpenAI o3 leaves ChatGPT on August 26, 2026 after a 90-day sunset, while GPT-4.5 leaves on June 27, 2026 after a 30-day sunset.
• Surface change: Canvas is no longer available in GPT-5.5 Instant or GPT-5.5 Thinking, with writing and coding moving into response blocks.
• Watch: OpenAI says this is ChatGPT only. The o3 and GPT-4.5 API paths are not changed by this notice.

OpenAI added two compact but operationally important notes to the ChatGPT release notes on May 28, 2026. The first note updates the response style and quality of GPT-5.5 Instant. The second note sets dates for removing OpenAI o3 and GPT-4.5 from ChatGPT. o3 gets a 90-day sunset and leaves ChatGPT on August 26, 2026. GPT-4.5 gets a 30-day sunset and leaves ChatGPT on June 27, 2026. OpenAI explicitly says the change applies only to ChatGPT and does not change API availability.

The same release note also changes the editing surface inside ChatGPT. Canvas is no longer available in GPT-5.5 Instant or GPT-5.5 Thinking. OpenAI says writing and coding functionality is now supported directly inside chat responses through writing blocks and code blocks. Paid users can keep limited Canvas access through legacy models for a transition period, but those legacy models now have dated exits from the ChatGPT picker.

The May 28 note is product cleanup, not just a model update

OpenAI had already begun rolling out GPT-5.5 Instant as the default ChatGPT model on May 5, 2026. That earlier note said GPT-5.5 Instant replaced GPT-5.3 Instant and improved accuracy, clarity, concision, image understanding, STEM answers, and web-search judgment. It also said Plus and Pro users would get better use of past chats, files, and connected Gmail context on the web. Paid users were told they could keep GPT-5.3 Instant through model configuration settings for three more months.

The May 28 entry does not announce a new default model. It says GPT-5.5 Instant has been tuned to respond in a more readable and natural style, with better pacing on practical-help tasks and fewer overly long or bullet-heavy answers. For builders, the larger product change sits in the next sentence: Canvas is removed from GPT-5.5 Instant and GPT-5.5 Thinking. A model quality update and a workspace simplification landed in the same public notice.

OpenAI's GPT-5.5 in ChatGPT Help Center article describes GPT-5.5 as three paths: Instant, Thinking, and Pro. Instant is the fast default for everyday work. Thinking is the deeper reasoning path for harder tasks. Pro is described as the highest-capability option for long-running workflows and the hardest tasks. ChatGPT can automatically switch from Instant to Thinking when a request is complex, and users can manually choose Thinking when they want the reasoning path.

That routing model means the user-facing model name is only part of the operating story. A person may start in Instant, ChatGPT may move the request into Thinking, and the result may appear as a writing block or code block instead of a separate Canvas document. Teams that document internal procedures as "run this in o3" or "edit this in Canvas" now have to update both the model choice and the surface where the artifact is maintained.

Canvas removal changes the unit of work

Canvas gave ChatGPT a separate place for longer writing, code, and iterative editing. OpenAI's May 28 wording says the underlying writing and coding features remain available through response blocks inside the chat. That is a narrow product sentence, but it changes the workflow boundary. The artifact now lives in the conversation unless the user moves it into an IDE, document editor, repository, or another tool.

For coding workflows, the distinction is practical. Canvas behaved like an adjacent editing surface where users could keep a larger file or draft visible while asking for targeted changes. A code-block-centered response leaves the output in the transcript and makes the handoff to the repository more explicit. A user asking ChatGPT to review a short function may not notice much. A team using Canvas for long refactoring notes, design documents, or mixed prose-and-code drafts will need a replacement surface before legacy access disappears.

There is also a documentation timing issue. The GPT-5.5 Help Center article says GPT-5.5 Thinking supports ChatGPT tools, and it separately says GPT-5.5 Instant supports current ChatGPT tools except Canvas. The newer May 28 release note says Canvas is no longer available in either GPT-5.5 Instant or GPT-5.5 Thinking. For this specific question, the release note is the more recent source. When OpenAI's product docs and release notes are updated at different speeds, builders should read the "Updated" timestamp and the release-note date together.

This is not only a preference change for writers. It affects repeatable work. If a team has a support runbook that says a human should paste a customer escalation into Canvas, ask o3 for a diagnosis, and revise the result in place, that runbook now has two dated dependencies. GPT-4.5 leaves ChatGPT on June 27, 2026. o3 leaves ChatGPT on August 26, 2026. Canvas remains available only through legacy models for a limited period. The API carve-out does not preserve a human workflow built around the ChatGPT UI.

o3 and GPT-4.5 are ChatGPT sunsets

OpenAI gave exact dates for both model removals. OpenAI o3 leaves ChatGPT on August 26, 2026 after a 90-day sunset. GPT-4.5 leaves ChatGPT on June 27, 2026 after a 30-day sunset. The release note describes these as paid-user models available through model settings, not as API deprecations.

The API distinction matters. OpenAI says the o3 and GPT-4.5 notice is ChatGPT only and that there are no API changes. Internal tools, evaluation harnesses, back-office automation, and production services that call o3 or GPT-4.5 by API are not directly broken by this announcement. The affected surface is ChatGPT: the model picker, legacy access, tools, Canvas availability, workspace policy, and any human process that depends on those controls.

This separation shows how OpenAI now operates two different delivery paths. The API path asks developers to choose model names in code, track pricing, and handle migration in their own release process. The ChatGPT path bundles model selection with automatic routing, memory, files, tools, connected apps, workspace settings, and role-based access. A model disappearing from ChatGPT can be urgent even when the same model name remains callable by API.

For engineering teams, the immediate audit is small and concrete. Search prompt libraries, internal docs, and enablement material for fixed references to o3, GPT-4.5, and Canvas. Separate workflows that call the API from workflows where people use ChatGPT directly. Put the two sunset dates next to the affected human process, not just next to the model name.

GPT-5.5 limits define the real operating boundary

OpenAI's GPT-5.5 documentation lists usage limits that shape the new default path. Free users can send up to 10 GPT-5.5 messages every five hours, then ChatGPT falls back to a mini version. Plus and Go users get up to 160 messages every three hours. Plus or Business users who manually select GPT-5.5 Thinking get up to 3,000 messages per week. Go users can enable Thinking through the plus icon in the chat entry box and send up to 10 Thinking messages every five hours.

Automatic routing has a different accounting rule. When GPT-5.5 Instant automatically routes a request to GPT-5.5 Thinking, OpenAI says that does not count against the manual Thinking weekly limit. Even after a user reaches the weekly manual Thinking cap, ChatGPT may still route to Thinking automatically when it decides the request requires it. In practice, "I selected Instant" and "the request actually used a reasoning path" can be different facts.

Context windows also differ by plan and mode. For GPT-5.5 Instant, Free gets 16K, Plus and Business get 32K, and Pro and Enterprise get 128K. For manually selected Thinking, paid tiers get 256K, while the Pro tier gets 400K. OpenAI breaks the Pro 400K figure into 272K input and 128K maximum output. Those numbers matter more than the model branding for long code reviews, multi-file document analysis, or large support investigations.

These limits also change how teams should write instructions for employees. A note that says "use GPT-5.5 for this review" is less precise than a note that says whether the task needs Instant, manually selected Thinking, automatic routing, a large context window, or a tool-enabled workspace. The old model-name shortcut becomes less useful when the product is routing requests and enforcing different limits behind the same front door.

Enterprise defaults are policy, not just capability

OpenAI's ChatGPT Enterprise and Edu Models and Limits article shows a different operating model from individual ChatGPT accounts. GPT-5.5 Instant access is disabled by default in ChatGPT Enterprise workspaces. An admin or owner has to enable it in workspace settings. GPT-5.5 is not available for ChatGPT for Healthcare workspaces. The same model family can be the default for a consumer user and an opt-in feature for an enterprise tenant.

Enterprise documentation also treats routing as a budget and governance control. Enterprise customers on a flexible pricing plan can configure Auto to use Thinking mini instead of Thinking for reasoning tasks. The document says Thinking mini usage does not consume credits. When GPT-5.5 Thinking is enabled in a workspace, members can manually choose it from the model picker. Those controls are closer to capacity planning than a simple model-quality preference.

Role-based access control adds another layer. OpenAI says routing logic has been updated so GPT-5.5 is used only when a user has access through RBAC. New chats use an equivalent model with access enabled. Existing chats can be blocked from sending new messages if an admin later disables access, returning an access denied error. Long-running project chats can therefore break because of a workspace policy change, even when the user still sees the historical transcript.

For enterprise administrators, the o3 and GPT-4.5 dates are only part of the migration. They also need to review whether GPT-5.5 Instant is enabled, whether Thinking is available, whether Auto routing is configured to use Thinking mini, and whether RBAC changes will strand existing chats. The product path is now a policy surface.

Legacy model access is a bridge, not a home

OpenAI's Legacy Model Access for Enterprise and Edu Users article lists GPT-4o, OpenAI o3, OpenAI o3 Pro, GPT-4.1, and GPT-4.5 as legacy models that can appear in the picker when legacy access is enabled. These models still consume credits and remain subject to limits. The framing is temporary: legacy access exists so teams can adapt workflows while GPT-5.5 becomes the default path.

That language is important because it changes how teams should treat the toggle. Turning on legacy access is not a migration plan. It is a way to buy time for users who still depend on an older model or Canvas behavior. Once o3 and GPT-4.5 have dated ChatGPT exits, a legacy toggle should trigger an inventory of the workflows using it.

The inventory should capture three things. First, identify internal instructions that name o3 or GPT-4.5 as if the ChatGPT picker were stable. Second, identify drafts, code review habits, and editing loops that depend on Canvas. Third, keep API workloads separate from ChatGPT workloads in the migration tracker. OpenAI preserved the API path in this notice, but the human-facing ChatGPT path is changing on a calendar.

The developer question is where the artifact lives

The May 28 update is easy to misread as another incremental model tuning note. The concrete changes are more operational. ChatGPT users start on GPT-5.5 Instant, can route into Thinking automatically or manually, and now receive writing and code artifacts inside response blocks rather than a GPT-5.5 Canvas surface. o3 and GPT-4.5 are dated for removal from ChatGPT. API users get time because this announcement does not deprecate those API models.

That leaves teams with a workflow question instead of a benchmark question. Short explanations, log summaries, and function-level fixes may fit GPT-5.5 Instant and response blocks. Large design docs, multi-file reviews, and long-context analysis depend on Thinking limits, plan tier, and workspace policy. Editing flows that used Canvas need a new home in a document tool, IDE, repository process, or the limited legacy window.

The bigger pattern is that ChatGPT and the API are now diverging as operational products. API users manage model names in code. ChatGPT users inherit a changing mix of model picker rules, automatic routing, Canvas removal, legacy toggles, Enterprise defaults, and RBAC. GPT-5.5 is not only a model transition. It is a cleanup of the screen where humans hand work to AI and take work back.