Strip the marketing away and the financial case for fine-tuning is one sentence: you are paying a training bill up front to stop sending the same 5,500 tokens of instructions and examples on every request for the rest of the model's life. Nothing else in the cost model matters as much. If your prompt is already short, fine-tuning cannot save you money — it can only make the output better, which is a different argument and one this calculator deliberately refuses to price.
The saving is prompt length, and only prompt length
Work the default scenario through. You send 6,000 prompt tokens today, most of it a system brief plus twenty few-shot examples. After tuning, the same behaviour needs 450 tokens. Output stays at 300 tokens either way, you run 150,000 requests a month, the training run and dataset preparation cost $600, and the platform charges a 50% premium on tuned inference. Base model: GPT-5.6 Luna at $0.20 input and $1.20 output per million tokens.
Prompting, per request: 6,000 ÷ 1,000,000 × $0.20 = $0.0012 input, plus 300 ÷ 1,000,000 × $1.20 = $0.00036 output. Total $0.00156, or $234 a month.
Tuned, per request: 450 ÷ 1,000,000 × $0.20 = $0.00009 input, plus the same $0.00036 output, giving $0.00045 — then multiplied by 1.5 for the premium, giving $0.000675. That is $101.25 a month.
The difference is $132.75 a month. Against $600 of training, payback lands at 4.5 months, or 677,966 requests. Year one nets $993. Real, worth having, and small enough that a single retraining cycle can erase it.
The premium applies to tokens you did not shorten
Here is the trap the arithmetic exposes. The 50% surcharge does not apply only to the input you cut — it applies to output too, and output is priced six times higher than input on this model. Your saving is capped by the input share of the bill, while the penalty scales with the whole thing.
Change one number to see it. Keep everything else identical and raise output from 300 tokens to 2,000 — a summariser, a drafting tool, anything that writes at length.
Prompting: $0.0012 + (2,000 ÷ 1,000,000 × $1.20 = $0.0024) = $0.0036 a request. Tuned: ($0.00009 + $0.0024) × 1.5 = $0.003735 a request.
The tuned model is now more expensive, by $0.000135 a request and $20.25 a month, despite a 93% shorter prompt. Nothing about the tuning got worse; the workload simply moved its weight to the side of the ledger that fine-tuning cannot touch. Before you budget a training run, check where your tokens actually sit. Output-heavy work is not a fine-tuning cost case at any prompt length. The LLM API cost calculator will show you the split in about thirty seconds.
Why platforms charge the premium at all
A tuned model is not the base model with a different flag set. The provider holds your adapter weights, routes your traffic to a deployment that can serve them, and loses some of the batching efficiency that makes shared inference cheap — thousands of customers hitting one set of weights is far more efficient than thousands hitting thousands of variants. The premium is what that lost efficiency costs, expressed as a per-token surcharge, and on some platforms as a flat hosting fee for keeping the deployment warm.
That is why the calculator has a separate monthly hosting field. If your provider bills for the endpoint whether or not you use it, a low-volume tuned model can be permanently underwater: $200 a month of hosting against $132.75 of token savings never comes right, no matter how many months you wait. Check which model your provider uses before you assume the premium field alone covers it.
Retraining resets the clock
The payback figure assumes the tuned model keeps earning indefinitely. It will not. A tuned model is frozen against a base model that keeps moving — new versions ship, the base you tuned from gets deprecated, your data distribution drifts as your product changes, and the behaviour you trained in slowly stops matching what you now want.
So the honest reading of a 4.5-month payback is not "profitable from month five". It is "profitable from month five until you retrain, at which point the $600 lands again". If you retrain quarterly, a 4.5-month payback never arrives. The rule that follows is blunt: if your payback period is longer than your retraining interval, the cost case does not close. You would be buying fine-tuning for quality, and you should say so out loud rather than dressing it up in a spreadsheet.
What prompt caching does to the argument
Caching attacks the same problem from the other direction, at zero training cost. GPT-5.6 Luna reads cached input at $0.02 per million against $0.20 standard. If that 6,000-token prompt is a stable prefix — a fixed system brief plus fixed examples, which is exactly what few-shot prompting produces — the cached version costs 6,000 ÷ 1,000,000 × $0.02 = $0.00012 instead of $0.0012.
Per request that is $0.00048 all in, against $0.000675 for the tuned model. Caching wins, needs no training run, no dataset, no premium, no retraining cycle, and can be switched off on a Tuesday afternoon if it disappoints. Run the prompt caching savings calculator against the same prompt before you commission a fine-tune; the mechanics are covered in prompt caching explained. On most workloads with a long stable prefix, caching captures the majority of the available saving for none of the commitment.
What this calculator will not tell you
There are good reasons to fine-tune that have nothing to do with tokens, and pricing them here would produce a false precision that helps nobody:
- Format reliability. A tuned model that emits valid JSON on 99.8% of calls instead of 98.5% removes a retry path, a validation layer and a class of pager alerts. That is worth money, but the amount depends entirely on what a malformed response costs you.
- Behaviour that will not fit in a prompt. Some tone, house style and domain judgement takes hundreds of examples to convey. Instructions describing it do not survive contact with edge cases.
- Latency. Shorter prompts prefill faster. On an interactive product a few hundred milliseconds may matter more than the $132.75.
- Small models punching up. A tuned small model matching an untuned mid-tier one on your narrow task changes which row of the price index you are shopping in, and that is a bigger lever than any of the above.
Put your real prompt lengths in, then check the year-one net figure against your honest retraining interval. If those two numbers argue with each other, the answer is caching.
Prices used by this calculator were last verified on 15 August 2026 from the vendors' own pricing pages. See the full price index for every figure and its source, the change log for what has moved recently, and the methodology for how the index is maintained and where its limits are.