Is OpenAI fine-tuning shutting down completely?

OpenAI is winding down creation of new jobs on its self-serve fine-tuning platform. Inference on completed fine-tuned models remains available until the relevant base model is deprecated.

Who can still create OpenAI fine-tuning jobs?

Organizations that had not previously used fine-tuning lost access on May 7, 2026. From July 2, organizations without fine-tuned-model inference in the previous 60 days also cannot create new jobs. Active existing customers retain job creation until January 6, 2027.

Should a new project use RAG instead of fine-tuning?

Use RAG when the main requirement is current or private knowledge with citations and fast document updates. Fine-tuning is a better fit for repeatable behavior, style, classification, extraction, or task-specific response patterns.

What should existing OpenAI fine-tuning customers do now?

Inventory models, training data, evaluation sets, prompts, owners, and base-model retirement dates. Then test a portable alternative before January 6, 2027 rather than waiting for the final job-creation cutoff.

Will existing fine-tuned models stop working on January 6, 2027?

No. OpenAI says inference continues until the underlying base model is deprecated. January 6 is the cutoff for active existing customers to create new fine-tuning jobs.

OpenAI Self-Serve Fine-Tuning Is Winding Down: What to Use Instead

OpenAI is winding down access to its self-serve fine-tuning platform. The change does not immediately disable completed fine-tuned models, but it does make the platform a poor foundation for new long-term customization projects. Teams should separate what actually requires training from what can be solved with prompting, retrieval, tools, or a more portable model platform.

The key distinction

New training-job access is being removed in stages. Inference on completed fine-tuned models continues until each underlying base model is deprecated.

Official availability timeline

Date	Change
May 7, 2026	Organizations that had not previously run fine-tuning could no longer create training jobs.
July 2, 2026	Organizations without fine-tuned-model inference during the previous 60 days could no longer create new jobs.
January 6, 2027	Active existing customers will no longer be able to create new fine-tuning jobs.
Base-model retirement	Inference on a completed fine-tuned model ends when its underlying base model is deprecated.

Source: OpenAI API Deprecations.

Do you actually need fine-tuning?

Fine-tuning is often proposed too early. Start by identifying the failure mode. If the model lacks current company knowledge, changing its weights is usually less practical than connecting it to a maintained source of truth. If the model understands the task but produces inconsistent structure or tone, better examples, structured outputs, tools, and evaluations may solve the problem without training.

Need	Start with	Why
Current or frequently changing knowledge	RAG / file retrieval	Documents can be updated without retraining.
Reliable JSON or workflow execution	Structured outputs, tools, and evals	The requirement is interface reliability, not new knowledge.
Consistent tone or response pattern	Prompt examples first; fine-tune if needed	Few-shot prompting is cheaper to test and reverse.
High-volume classification or extraction	Fine-tuned smaller model	Training can improve consistency and unit economics at scale.
Full deployment and weight control	Open-weight model platform	Supports self-hosting, portable artifacts, and deeper customization.

Four replacement paths

1. Prompting, structured outputs, and evaluations

Best when the underlying model can already perform the task. Build a representative evaluation set, improve instructions and examples, and enforce machine-readable output before introducing training infrastructure.

2. Retrieval-augmented generation

Best for policies, product documentation, customer records, research, and other knowledge that changes. Keep permissions, document freshness, retrieval quality, and citations observable.

3. Managed training on another cloud or model platform

Best when the team wants hosted infrastructure but needs a longer-lived training path. Compare supported base models, artifact portability, evaluation tools, regions, data retention, and exit costs.

4. Open-weight fine-tuning or self-hosting

Best when model ownership, privacy, custom deployment, or vendor independence matters enough to justify GPU infrastructure, security, serving, monitoring, and upgrade responsibilities.

Migration checklist for existing customers

Inventory every fine-tuned model, base model, owner, endpoint, and production consumer.
Export and document training and validation datasets while access and institutional knowledge are available.
Preserve the evaluation set separately from training data so alternatives can be compared fairly.
Record current quality, latency, throughput, failure rate, and total cost as a migration baseline.
Test prompting and RAG before selecting another training platform.
If training remains necessary, compare hosted and open-weight options with the same evaluation set.
Map each fine-tuned model to its base-model retirement date and create an owner-backed exit deadline.
Deploy the replacement gradually with logging, human review, and a rollback path.

AILinkBase recommendation

Do not start a new product whose core dependency is the ability to create OpenAI self-serve fine-tuning jobs. Existing customers do not need to switch off working models immediately, but they should treat January 6, 2027—and the retirement date of each base model—as hard planning constraints. The safest architecture keeps prompts, evaluation data, retrieval content, and routing logic portable across model providers.