What a fine-tune costs against the base model you gave up

Fine-tuning is usually costed as the training run, which is the small part and the only part that arrives as a single obvious charge. Two larger costs sit either side of it. The first is inference: a tuned model bills at its own rates for as long as you serve it, so a rate difference of cents per million compounds across every call for the life of the deployment. The second is the model you gave up. On OpenAI's published pricing, read 2026-08-05, every model priced for fine-tuning is prior generation, and the GPT-5 family appears nowhere in that section. So committing to a fine-tune means committing to a gpt-4.1, gpt-4o or o4-mini generation model, and the base rates you're comparing against keep moving without you. The comparison also refuses to resolve in general, because it depends on the shape of your traffic rather than on which model is better. Fine-tuned gpt-4.1 is cheaper than base gpt-5.4 on output and dearer on input, so there's an exact ratio where they cross. Which side you land on is a property of your own traffic, and it also depends on whether the fine-tune changes your prompts or only the model running them. Those two readings of one decision disagree.

• Your fine-tuning business case counts the training run and stops.
• Nobody compared the tuned model's inference rate to the base model's.
• You fine-tuned a model a generation behind and never revisited it.
• You picked o4-mini, whose training bills by the hour, and budgeted it from a token count.
• You can't say whether your traffic is input-heavy or output-heavy.

1. Decide first whether the fine-tune would change your prompts or only the model behind them.
2. If only the model, work out your input to output ratio and see which side of 6:1 you sit on.
3. Put the tuned model's inference rate next to the frontier model you would otherwise run.
4. Multiply the rate difference by a year of your call volume before deciding.
5. Check whether the model you fine-tuned is still the one you would choose today.

Illustrative example

Fine-tuned gpt-4.1 bills $3.00 input and $12.00 output per million. Base gpt-5.4 bills $2.50 and $15.00. Both are standard rates read off OpenAI's pricing page on 2026-08-05, without the data-sharing inference discount that page offers. The gpt-5.4 rates match this price book exactly and the fine-tuned rates sit outside it entirely. The tuned model is dearer on input and cheaper on output. Token shapes below are modelled.

The 6:1 crossover answers one question: swap the model, keep the prompts. It doesn't answer the question people actually ask, which is whether to fine-tune so they can stop sending the context. Those are different comparisons and the same two rate cards give opposite answers, $0.0720 against $0.0650 one way and $0.0180 against $0.0650 the other. Work out which question you're asking before you use either number, because the arithmetic is easy and picking the wrong comparison is the mistake that actually gets made.

Both rate cards are public and the crossover is arithmetic anyone can do. What decides it: your own input to output ratio, which is a property of your traffic that no pricing page contains. It also moves: add retrieval and the ratio climbs, tighten a prompt and it falls, and a fine-tune that made sense at 3:1 quietly stops making sense at 8:1 without anything announcing it. Culpa prices every call from a versioned price book in exact decimal and keeps the token shape alongside the cost, so the ratio is a number you have rather than one you assume, and the comparison can be re-run against the traffic you actually served rather than the traffic the business case imagined.

• Model Upgrade Cost Increase: Same Workload, New Rate
• LLM Model Mix Drift: The Bill Rises With No Code Change
• Model Mix Cost Teardown: The Cheap Model Did More Work
• Model Switch Cost Calculator: Price the Move Before You Make It