Introduction

Today, June 27, 2026, GPT-4.5 is gone from ChatGPT. Not deprecated with a long runway, not gently wound down over twelve months while users adjusted — gone, following a 30-day sunset period that OpenAI announced quietly in its ChatGPT release notes, without a dedicated blog post, without a press release, and without the kind of advance notice that most businesses reasonably expect when a tool they have integrated into their products is being discontinued. GPT-4.5 was released in February 2026 as OpenAI's most knowledgeable model at launch. It survived four months. The retirement comes on the same day it was scheduled, with OpenAI noting in its release notes that "OpenAI o3 will be retired from ChatGPT on August 26, 2026 following a 90-day sunset period, and GPT-4.5 will be retired from ChatGPT on June 27, 2026 following a 30-day sunset period."

Four months from launch to retirement. That timeline is fast even by OpenAI's standards, and OpenAI's standards are already faster than the industry was prepared for. The retirement is not a product failure — it reflects the pace at which OpenAI's own model lineup is advancing, with GPT-5 and its successors having fully superseded GPT-4.5 on every benchmark that mattered within months of its launch. But the business implication of that pace is the story that the retirement notice itself does not tell. When the most capable model from a major AI provider can be created, deployed, used by millions, and retired faster than most companies conduct an annual performance review cycle, the companies building products on those models need a fundamentally different infrastructure strategy than the one that most of them currently have.

What GPT-4.5 Was and Why People Valued It

GPT-4.5 launched in February 2026 as a large, expensive, conversationally capable model optimised for broad world knowledge and smooth, natural dialogue. It was not designed to be the fastest or cheapest model in OpenAI's portfolio — GPT-4o and its successors handled that role — and it was not the deepest reasoning model, a distinction that belonged to the o-series. What GPT-4.5 was designed for was the quality of its conversational texture: the warmth of its responses, the breadth of its contextual knowledge, and the fluency with which it integrated both into natural language interactions.

Developers and power users who chose GPT-4.5 over other contemporaneous options frequently cited exactly those qualities. A vocal minority of users pushed back against the retirement on social media, with one widely cited comment on X summarising the sentiment: "To this day, 4.5 is the best writing model. o3 was a native pure reasoning model. 5 series still doesn't match what those two had." That response reflects a pattern that OpenAI has now encountered multiple times during its model retirement cycles: the successor model may score higher on aggregate benchmarks, but specific capabilities that some users valued — a conversational style, a particular reasoning approach, a preferred personality — do not always transfer cleanly across model generations.

This matters for business users who built products specifically around GPT-4.5's characteristics. A company that chose GPT-4.5 for a customer-facing writing assistant because its output matched their brand voice more consistently than GPT-5's default style now faces a migration that is not simply a model string swap in an API call. It is a re-evaluation of whether the replacement model produces the same outputs their customers expect, which requires regression testing, prompt adjustment, and potentially user research before production traffic can safely move.

What Replaced It and How the Retirement Actually Works

Before examining what to migrate to, it is important to clarify what today's retirement actually covers — because the framing in OpenAI's announcement contains a distinction that has significant implications for developers with active API integrations.

The GPT-4.5 retirement announced for June 27, 2026, applies to ChatGPT only. OpenAI's release notes were explicit: "These changes apply to ChatGPT only; there are no changes to the API." This is consistent with how OpenAI has handled several earlier retirements in the current deprecation cycle: models are first removed from the ChatGPT interface, where usage is lower and switching friction for end users is manageable, before API deprecation follows on a separate, typically later schedule. Developers using GPT-4.5 through the OpenAI API on the model string "gpt-4.5-preview" should consult the OpenAI deprecation documentation directly for the API shutdown date, which is separate from today's ChatGPT removal. OpenAI's general policy for generally available models is a minimum six-month notice period before API deprecation; for specialised variants and preview models, that window can be as short as three months or two weeks respectively.

For ChatGPT users who have been accessing GPT-4.5 through the model picker, the replacement in the ChatGPT interface is effectively GPT-5.5 Instant — the model OpenAI released on April 23, 2026, and has been upgrading incrementally through May and June to produce shorter, less structured responses with more natural conversational flow. For API developers whose integrations have been tested against GPT-4.5's output characteristics, OpenAI's recommended migration path is to GPT-5 or GPT-5.3-Codex, both of which OpenAI states are faster, cheaper per token, and score higher on most benchmarks than GPT-4.5 did at launch. The migration itself is mechanically straightforward: update the model parameter in the API call. The non-mechanical part — validating that the replacement model produces acceptable outputs for every use case that GPT-4.5 was handling — is where the actual migration work lives.

Why OpenAI Is Accelerating Its Deprecation Cycle

The pace of OpenAI's model retirement schedule in 2026 is not random. It is the direct consequence of a model development velocity that has no precedent in the history of commercial software.

The retirement timeline that played out in 2026 before today's GPT-4.5 removal illustrates the speed clearly. GPT-4o was retired from ChatGPT on February 13, 2026. GPT-5.1 models were retired on March 11, 2026. GPT-5.2 became unavailable in ChatGPT as of June 12, 2026. GPT-4.5 retires today. OpenAI o3 is scheduled for retirement on August 26. Within a seven-month window, the most widely used AI product in the world has cycled through multiple complete model generations, retiring each one as its successor proved demonstrably better in aggregate across the user population.

OpenAI has been honest about the logic driving this. When it retired GPT-4o in February, the announcement noted that only 0.1% of users were still choosing GPT-4o each day, with the vast majority having migrated to GPT-5.2 on their own. Maintaining a model that 0.1% of users choose requires compute resources, engineering maintenance, and safety monitoring that produce no proportionate benefit. From OpenAI's perspective, retiring low-usage models is straightforward resource allocation. From the perspective of the 0.1% of users who were still choosing GPT-4o — or who had built production systems around it — the retirement is not a resource allocation decision. It is a product discontinuation that affects their live applications.

The IPO context gives the acceleration an additional dimension. OpenAI filed its confidential S-1 with the SEC in early June 2026. Preparing for a public offering creates incentives to simplify the product portfolio, reduce operational complexity, and present investors with a cleaner picture of which models carry forward revenue and which represent legacy overhead. Model retirements that were always going to happen eventually may be happening slightly faster than they otherwise would, because a leaner model portfolio is easier to explain to prospective shareholders than a sprawling lineup of overlapping products with unclear differentiation.

The Risk This Creates for Businesses Building on a Single Provider

The GPT-4.5 retirement is a four-month data point in a longer pattern, and the pattern is the business risk. When a model that shipped in February is retired in June, and when that retirement is not an isolated event but the fourth in a sequence of retirements spanning a single year, the implicit contract between an AI model provider and the businesses building on its platform is being renegotiated at a pace that those businesses did not price in when they made their initial build decisions.

The companies most exposed to this risk are those that made two connected decisions simultaneously: they built on a single AI provider, and they built around a specific model rather than a model family alias. A product that calls "gpt-4.5-preview" explicitly in its API integration is tightly coupled to a model version that has now been retired from ChatGPT and will eventually be retired from the API. A product that calls "gpt-5" and relies on OpenAI's automatic routing to the current best version of that model family has more resilience, but it also has less control over output behaviour when OpenAI silently updates what "gpt-5" points to.

Neither approach is fully safe. The explicit model version approach breaks when the model is retired. The floating alias approach breaks when the model it points to changes behaviour in ways that affect the application's outputs without any code change on the developer's side. OpenAI has acknowledged this tension directly in its API documentation, noting that the gap between a ChatGPT interface retirement and an API deprecation is specifically designed to give developers time to evaluate replacements and complete migrations before the API endpoint closes.

The downstream effect on third-party tooling is also real. LangChain, LlamaIndex, and other orchestration frameworks that integrate with OpenAI's API shift their defaults as upstream models are deprecated. A developer who has not updated their dependencies may find that a library upgrade silently changes which model is being called, which changes the outputs, which changes the user experience, without any direct notification from OpenAI.

The Correct Long-Term Strategy for Any Company Building on AI in 2026

The GPT-4.5 retirement, taken alongside the broader deprecation pattern, points clearly toward what a sustainable AI infrastructure strategy looks like — and what it does not.

What it does not look like is a tight coupling between a production application and any specific named model version from any provider. The life expectancy of a specific model version in 2026 is measured in months, not years. A development cycle that includes a production deployment, integration testing, user acceptance testing, and a staged rollout across a user base can easily consume six months of the window between a model's launch and its deprecation. Companies that complete that cycle and then allow the resulting integration to remain static — because it is working and migration is expensive — are building on a substrate that is being withdrawn from underneath them on a timeline they do not control.

The correct architecture has several properties that the current generation of single-provider, version-specific integrations lacks. The first is provider abstraction: a middleware layer between the application and the model provider, so that switching from one model or provider to another requires a change in the middleware configuration rather than a change in the application code. A model gateway, an orchestration framework with provider-agnostic routing, or a careful use of floating model aliases in the API layer all achieve versions of this property. The second is multi-provider testing as a standard part of the release cycle: regularly validating that core application functionality produces acceptable outputs not just on the primary model but on the leading alternative models, so that provider migration is a well-practised operational capability rather than an emergency response to a deprecation notice.

The third property is monitoring for output drift: systematic measurement of model output characteristics in production, so that behaviour changes caused by silent model updates, floating alias redirects, or provider-side updates are detected and assessed quickly rather than discovered through user complaints. The fourth is a written migration playbook specific to each AI integration in the product, maintained and tested on a cadence aligned with the provider's deprecation schedule rather than on the crisis timeline that a 30-day sunset period imposes.

None of this is novel advice. It describes the same resilience engineering principles that good software teams apply to any critical dependency. What the GPT-4.5 retirement makes concrete is that those principles apply to AI model dependencies with a deprecation frequency that is significantly higher than most critical software dependencies, and that the businesses which built their AI integrations during the 2023-2024 period of relative model stability are now facing that frequency for the first time.

Conclusion

GPT-4.5 launched on February 28, 2026. It retired today, June 27, 2026. The interval is 119 days, which is the shortest model lifecycle from launch to ChatGPT retirement in OpenAI's history. That record will almost certainly be broken by a successor model in the current deprecation cycle, because the model development velocity that made GPT-4.5 obsolete in four months has not slowed.

For everyday ChatGPT users, today's retirement is a minor inconvenience — the model picker now defaults to GPT-5.5 Instant, which is meaningfully better on most tasks, and most users have already migrated without noticing. For developers and businesses with active integrations built around GPT-4.5's specific characteristics, today is a forcing function: test the replacement, update the integration, and then build the infrastructure that prevents the next deprecation from being a forcing function too.

The broader message of the GPT-4.5 retirement is not about this specific model. It is about the nature of AI infrastructure in 2026. The models are getting better faster than at any previous point in computing history. That improvement comes at a cost: no model version is permanent, no provider guarantee is indefinite, and no architecture that treats a specific named model as a stable dependency is robust to the deprecation cycle that makes those improvements possible. The businesses that thrive in this environment are those that treat AI model dependencies the way experienced engineers treat every other fast-moving dependency: with abstraction, with testing, with monitoring, and with a migration playbook they practised before they needed it.