OpenAI is killing products within months and folding their features back into the core

OpenAI has ended or begun withdrawing four standalone products within a matter of months. Ethan Mollick puts their lifespans at just 6 to 9 months. The pattern matters more than the list: models are changing faster than the product shells built around them.

Four short lives expose the speed of the reversal

Mollick says Sora lasted 6 months, Agent Builder and Group Chat 8 months each, and Atlas 9 months. OpenAI published a Sora shutdown guide covering content exports, data deletion, refunds and a planned API closure. Agent Builder also received a deprecation notice after launching as part of AgentKit in October 2025.

The figures in Mollick's post are his tally of product lifespans, not an OpenAI metric. His broader point still holds: the company is not merely removing failed experiments. It often retires a standalone interface after its own models or core products absorb the same job in another form.

Customers are buying a bet on the wrapper, not just the capability

Moving a feature into ChatGPT may be convenient for an individual. For a team that built workflows, training or integrations around a product, an eight-month lifespan means another migration. The material risk is therefore not only model quality, but the durability of a particular product and API.

OpenAI is also demonstrating a practical feature of exponential development: today's product decision can age before a large company completes procurement and rollout. The vendor with the strongest models may still have the least predictable product map.

Fast shutdowns can be discipline or the bill for a bad forecast

Killing a dead end is healthier than maintaining it for years. Four short cycles still raise the price of trust. Export and transition guidance can preserve data, but it cannot return the time spent on integrations or guarantee that the replacement covers the same use cases.

Mollick is offering an interpretation rather than an audit of every decision. Without internal data, outsiders cannot separate weak adoption, cost, security concerns and simple consolidation into ChatGPT. Each closure may have a different cause.

Migration paths and API stability will decide whether users stay

The next useful signal is whether OpenAI provides compatible replacements, clear deadlines and exports that preserve work. Enterprise buyers should also watch whether similarly short cycles spread to APIs and agent tooling.

The sensible response is not to freeze adoption. It is to design for replacing both vendor and interface. A product that lasts 6 months can be a useful experiment, but without an exit route it should not become a load-bearing wall.

Lilith's verdict

OpenAI can discard its own map before a customer has time to hang it on the wall. Anyone building operations on it needs a marked emergency exit beside the integration.