Fine-Tunable Models | When and Why to Fine-Tune

Not every OpenAI model supports fine-tuning, and the ones that do support different training methods. Choosing the wrong model for your task is like buying a sports car for off-road driving -- technically possible, but you are fighting the tool instead of using it.

The Current Lineup

OpenAI offers fine-tuning across three tiers of capability, each with distinct cost profiles and training method support.

Three variants: gpt-4.1 (full power, highest cost), gpt-4.1-mini (balanced capability and cost), and gpt-4.1-nano (fastest, cheapest, surprisingly capable for focused tasks). All three support Supervised Fine-Tuning (SFT). The 4.1 family excels at instruction following, long-context understanding, and structured output generation. Choose gpt-4.1 when quality is paramount and cost is secondary. Choose gpt-4.1-mini for production workloads where you need strong performance at reasonable cost. Choose gpt-4.1-nano for high-volume, focused tasks like classification or extraction where speed and cost matter more than nuanced reasoning.

Model Selection Matrix

Use Case	Recommended Model	Training Method	Why
Customer support tone	gpt-4.1-mini	SFT	Good balance of quality and cost
Legal document analysis	gpt-4.1	SFT	Needs maximum reasoning capability
High-volume classification	gpt-4.1-nano	SFT	Speed and cost at scale
Brand image recognition	gpt-4o-2024-08-06	Vision SFT	Only option for visual fine-tuning
Complex math reasoning	o4-mini	RFT	Learns to reason, not just pattern-match
Code style enforcement	gpt-4.1-mini	SFT	Code understanding + affordable

Pricing Implications

Fine-tuning costs have two components: training and inference. Both vary significantly by model.

Training costs are charged per token in your training dataset. Larger models cost more to train -- gpt-4.1 training tokens cost roughly 4x what gpt-4.1-nano tokens cost. A dataset of 1,000 examples at ~500 tokens each (500K total tokens) might cost:

gpt-4.1-nano: ~$2-5
gpt-4.1-mini: ~$5-15
gpt-4.1: ~$15-50

Inference costs for fine-tuned models are higher than base model inference. You are paying a premium for the customized behavior. The exact multiplier varies, but expect roughly 1.5-3x the base model's per-token rate.

The Hidden Cost: Data Preparation

Training compute is rarely the largest cost. A 500-example dataset might cost $10 to train but $5,000 in expert time to curate. Domain experts writing, reviewing, and validating training examples is almost always the bottleneck -- both in time and in money. Budget accordingly.

Choosing Your First Fine-Tuning Target

If you are new to fine-tuning, start with gpt-4.1-mini. Here is why:

Cost forgiveness: Training mistakes are cheaper to repeat
Fast iteration: Smaller models train faster, so you can run more experiments
Strong baseline: Mini is capable enough that fine-tuning improvements are meaningful
Production viable: Many production workloads run on mini-class models

Once you have validated your data pipeline and evaluation methodology on mini, you can scale up to gpt-4.1 if quality demands it -- or scale down to nano if cost demands it.

Start with gpt-4.1-mini for your first fine-tuning experiment. Validate your data quality, training pipeline, and evaluation methodology before scaling to more expensive models.

Don't

Jump straight to gpt-4.1 for your first attempt. If your training data has quality issues -- and it almost certainly does on the first try -- you will burn through budget learning lessons you could have learned on a cheaper model.

Base Model Updates and Version Pinning

OpenAI periodically updates base models. When the base model changes, your fine-tuned model does not automatically update -- it remains on the snapshot it was trained on. This means:

Your fine-tuned model will not benefit from base model improvements unless you retrain
Your fine-tuned model will not break from base model changes either
You should version-pin your training data and evaluation sets so you can retrain deterministically

Concept Card

Method Compatibility Summary

Model	SFT	DPO	RFT	Vision FT
gpt-4.1	Yes	Yes	No	No
gpt-4.1-mini	Yes	Yes	No	No
gpt-4.1-nano	Yes	No	No	No
gpt-4o-2024-08-06	Yes	No	No	Yes
o4-mini	No	No	Yes	No

Notice that DPO (Direct Preference Optimization) is only available on gpt-4.1 and gpt-4.1-mini. If preference alignment is your goal, these are your only options. We will cover DPO in detail in Chapter 3.

beginner

You are building a customer support classifier that needs to categorize incoming tickets into 8 categories. The system handles 10,000 tickets per day. Write down: (1) Which model would you choose for fine-tuning and why? (2) What training method (SFT, DPO, or RFT) fits this task? (3) Estimate the monthly inference cost difference between your chosen model and gpt-4.1 at 10,000 calls/day with ~300 tokens per response.