When to fine-tune instead of prompt

Fine-tuning is often reached for too early. It's slower to iterate, harder to debug, and locks you into a model version. Here's when it's actually worth it.

When prompting is enough

• The task is well within the base model's capability, just needs steering
• You can express the desired behavior in a system prompt + a handful of examples
• Requirements still change week to week

When fine-tuning wins

• You need a consistent output format at very high volume, where token cost from few-shot examples adds up
• The task requires domain vocabulary or style the base model doesn't have
• Latency matters and a smaller fine-tuned model can replace a larger prompted one

The cost math

Run the numbers before committing: training cost + iteration time vs. the per-request token savings from dropping few-shot examples. If you're not saving enough tokens per request to offset the training and maintenance cost within a few months, stick with prompting.