OpenAI’s ‘Strawberry’ Warnings Threatened Access, Not an Immediate Ban
OpenAI did warn some o1-preview users in September 2024 after prompts appeared to seek the model’s hidden reasoning. The messages threatened loss of access after further violations; they did not announce an immediate, platform-wide ban on everyone who merely asked how the model reached an answer.
That distinction matters even more now. As of August 13, 2026, o1-preview is no longer a current ChatGPT model, and the public interface has moved through several generations of reasoning systems. OpenAI continues to expose controls and summaries around reasoning rather than an unfiltered internal transcript, but its current public documentation does not say that an ordinary question about reasoning automatically triggers account termination.
What the 2024 warnings actually said
The controversy began days after OpenAI released o1-preview and o1-mini on September 12, 2024. These models were widely associated with the internal codename “Strawberry” and were designed to spend additional computation on a problem before returning an answer.
Users then posted screenshots of emails saying particular requests had been flagged as attempts to circumvent safeguards. Contemporary reporting by Ars Technica documented warnings that additional violations could lead to loss of access to “GPT-4o with Reasoning,” an internal label associated with o1 at the time. The report also described inconsistent user accounts about triggers, ranging from requests for a “reasoning trace” to broader questions about reasoning.
The available evidence therefore supports a narrower conclusion than the viral headline suggested. OpenAI was detecting prompts it classified as safeguard-circumvention attempts and warning affected users about future access consequences. The reporting did not establish that all questions containing the word “reasoning” were prohibited, nor did it document a mass removal of accounts.
Why users could see a summary but not the raw trace
o1-preview generated intermediate reasoning before composing its final response, but the text shown in ChatGPT was not a direct window into every internal step. OpenAI presented a model-generated summary, creating an important difference between asking the system to explain an answer and trying to extract the concealed trace itself.
In its original technical explanation, OpenAI’s o1 release research said it had decided not to display raw chains of thought after weighing user experience, competitive advantage and the possibility of monitoring those chains for model misbehavior. The company argued that an unsupervised internal trace could be useful to safety researchers inside OpenAI while also containing material unsuitable for direct presentation to users.
This was not a claim that users should accept answers without explanation. A final answer can include calculations, evidence, assumptions or a concise derivation written for the reader. What remained unavailable was the model’s complete hidden scratch work, which may differ from a deliberate explanation produced in the answer.
The model at the center of the dispute is now obsolete
Calling the episode a current “Strawberry ban” obscures how much the product has changed. o1-preview was an early preview, later replaced by the production o1 model and then by newer reasoning families. Its warning emails describe a specific launch-period enforcement episode, not the names or controls visible in ChatGPT today.
The current ChatGPT release notes show that by June 2026 the consumer model picker used reasoning-effort choices such as Medium, High and Extra High, with GPT-5.5 models remaining available after GPT-5.2 was retired. Those notes describe selectable thinking effort and, for newer models, user-facing plans that can help people steer a response while it is running. They do not describe access to an unfiltered internal chain of thought.
The enduring issue is therefore broader than o1. OpenAI has made reasoning more visible as a product experience—through progress indicators, adjustable effort, plans and explanations—without treating the underlying private computation as a user-readable log.
What users can safely ask for
A request for a useful explanation is not the same as an extraction attack. Users can ask a model to show calculations, identify assumptions, cite evidence, compare alternatives, check an answer against stated criteria or provide a concise derivation that another person can audit.
The important boundary is intent and technique. Prompts designed to reveal hidden instructions, reproduce private system material or bypass safeguards may be treated as attempts to circumvent protections. Simply asking “Why is this answer correct?” does not, on the public evidence reviewed here, establish the same behavior.
For creators, researchers and developers, the practical limitation is that a polished explanation should not be mistaken for a forensic record of the computation that produced the response. It can still be evaluated for factual accuracy and logical completeness, but it cannot prove that every hidden intermediate step was faithfully disclosed.
What remains unresolved
OpenAI has not published a detailed retrospective identifying how many warning emails were sent in 2024, how many users subsequently lost o1 access or whether particular harmless phrases caused false positives. Without those figures, claims of either a sweeping purge or a harmless one-off glitch go beyond the public record.
The confirmed picture is more restrained: warning emails were reported, further violations were said to risk model access, and OpenAI deliberately withheld o1’s raw reasoning. The specific preview model has since disappeared from current ChatGPT, but the transparency conflict it exposed remains: users can request evidence and an intelligible explanation, while the internal reasoning trace stays under the provider’s control.
Also read:
Subscribe to our newsletter
Get the latest Web3, AI, and crypto news delivered straight to your inbox.