OpenAI fine-tuning now closes new jobs for 60-day inactive orgs

OpenAI has narrowed self-serve fine-tuning again, and the operational detail matters: recent usage now determines whether an organization can create new jobs. The OpenAI API deprecations page says that as of July 2, 2026, organizations that have not run inference on a fine-tuned model in the past 60 days can no longer create fine-tuning jobs. That extends an earlier May 7 restriction on organizations that had not previously run fine-tuning.

Two dates define the migration window. July 2, 2026 is the cutoff for organizations without recent fine-tuned model inference. January 6, 2027 is the broader deadline, when active existing customers will also lose the ability to create new fine-tuning jobs. OpenAI says inference on fine-tuned models will remain available until the underlying base model is deprecated, so this is not an immediate shutdown of already deployed fine-tuned systems.

The practical impact lands on teams that treated fine-tuning as an occasional workflow. A model may still exist, but if the organization has not used a fine-tuned model for inference within the 60-day window, it can lose the ability to start new training jobs. That turns usage history into part of the deployment surface, alongside model IDs, base-model deprecation dates, and application code.

For production teams, the work is straightforward but time-sensitive. Identify which applications depend on fine-tuned models, verify whether those models have recent inference traffic, and decide whether any retraining needs to happen before January 6, 2027. Teams using fine-tuning for domain-specific extraction, classification, tone control, or structured output should also review whether the same behavior can be maintained with newer base models, retrieval, tools, or prompt changes.

The larger signal is that OpenAI is reducing the long tail of self-serve training paths while keeping inference alive until base models age out. Fine-tuning is not disappearing overnight, but the default path for new customization is being constrained. Developers who rely on it need a calendar, not just a fallback model name.