What Fine-Tuning Costs Before and After Training

Fine-tuning has two bills: the training job and every request sent to the tuned model afterward. For example, Google Cloud lists Gemini 2.5 Flash supervised tuning at $5 per million training tokens and standard text inference at $0.30 per million input tokens and $2.50 per million output tokens. A job that processes one million training tokens would therefore cost $5 in training charges, before evaluation work and future requests. Those figures come from Google Cloud’s pricing page; they are not a quote for a complete project. The important question is whether the tuned model can do the same job with fewer billed tokens, better quality, or both.

What goes into the total bill

A useful budget separates provider fees from the work your team must supply. Count the examples you collect, clean and label; the test cases you hold out; the training run; any repeat runs; evaluation calls; and the requests made after deployment. The API invoice will not tell you what your staff time costs. Put that amount in your worksheet rather than treating it as free.

Training billing also differs by service. Google Cloud says its supervised Gemini training tokens equal the tokens in the training dataset multiplied by the number of epochs. An epoch is one pass over the training data. Together AI says its charge also includes optional validation tokens multiplied by the number of evaluations. Thus, “one million tokens in my file” is not automatically one million billable training tokens.

Serving can change the economics more than a small training fee. For the Gemini 2.5 models used here, Google’s pricing page lists the same prediction rates for tuned and base models, and its deployment guide says the tuned adapter has no endpoint charge. The deployment guide states price parity broadly, but the pricing page specifies a 1.5× base-model prediction rate for Gemini 3 and later tuned models in non-global regions; do not extend the Gemini 2.5 calculation to those models. Together AI instead says a dedicated endpoint for a fine-tuned model incurs continuous hosting charges, including idle time. Check the deployment method before comparing prices: a low training rate is not a low total cost if a GPU endpoint must run all month. (Together fine-tuning billing guide)

Published rates for three fine-tuning options

These are examples of services that list fine-tuning rates, not a ranking of model quality. Prices are US dollars. Google inference figures below are standard text rates for requests within its listed 200,000-token input tier; other modalities, priority or batch processing can cost differently. Together’s trained model needs a deployment quote for its dedicated endpoint, so its training figure cannot be treated as the whole bill.

Service and model	Training rate	Serving rate or fixed charge	What to check
Google Cloud Gemini 2.5 Flash-Lite, supervised tuning	$1.50 per 1M training tokens	$0.10 input and $0.40 output per 1M text tokens; Google says no tuned-endpoint charge	Quality on your task and any paid evaluations
Google Cloud Gemini 2.5 Flash, supervised tuning	$5 per 1M training tokens	$0.30 input and $2.50 output per 1M text tokens; Google says no tuned-endpoint charge	Whether shorter prompts offset training cost
Together AI Llama 3.1 8B, supervised LoRA	$0.34 per 1M processed tokens, $4 minimum per job	Dedicated endpoint hosting is charged continuously; select hardware to get a rate	Validation tokens, hardware, idle hours and deployment availability

The Gemini model support list is in Google’s tuning documentation. Its pricing table supplies the training and inference rates. Together’s pricing table supplies the Llama 3.1 8B LoRA rate and minimum; its billing guide explains validation and hosting. Confirm current prices in your own account before starting a job, because model availability and rates can change.

OpenAI is a useful caution against copying old price tables. Its current API pricing page says its fine-tuning platform is winding down and is no longer accessible to new users. An older article showing an OpenAI tuning rate does not establish that a new account can purchase that job today.

Calculate the training run

Start with the actual token count in the training set, not its file size in bytes. Multiply it by epochs, add any provider-billed validation tokens, and apply the model’s training rate:

Training charge = max(job minimum, [(training tokens × epochs) + (validation tokens × evaluations)] ÷ 1,000,000 × rate per million)

Use the minimum only when the provider lists one. For Google supervised Gemini tuning, its documentation specifies the training-set tokens times epochs; do not assume the Together validation rule applies to Google. Google also says an automatically configured Gen AI evaluation is billed separately as batch prediction. Budget those evaluation requests if you enable them. (Google tuning guide)

Here is an illustrative Gemini 2.5 Flash run, with invented workload inputs rather than a claimed test. Suppose a prepared dataset contains 200,000 tokens and the job uses five epochs. That processes 200,000 × 5 = 1,000,000 training tokens. At Google’s listed $5 per million, the training charge is $5. A second run with the same settings would add another $5. Human preparation, automatic evaluations and inference are outside this number.

The same one-million-token processed total would mathematically produce $0.34 at Together’s listed Llama 3.1 8B LoRA rate, but the posted $4 minimum makes the job charge $4 before any endpoint hosting. This is a different model and deployment setup; it is not a claim that the two produce equivalent results. Together says its dashboard shows the final token count and price after tokenization. (Together pricing; Together billing guide)

Add the monthly serving bill

A fine-tuned model is useful only if it meets the same evaluation target as the baseline. Compare both on one held-out set. Then measure the mean input and output tokens per request in a representative workload. Count all billed output tokens, including reasoning tokens, rather than only the visible answer. Google’s pricing table includes response and reasoning in text output. For the illustrative Gemini 2.5 Flash calculation below, assume thinking is disabled for both versions and the 200 output tokens are the complete billed output; Google’s tuning guide recommends a thinking budget of 0 for tuned Gemini 2.5 tasks. The simple text-only formula is:

Monthly serving cost = requests × [(mean input tokens × input rate) + (mean output tokens × output rate)] ÷ 1,000,000 + fixed monthly hosting

Keep input and output separate. They usually have different rates. Also keep retries, tool calls, cached-token rates and batch discounts separate if your application actually uses them. Our API price comparison explains ordinary token billing; the API cost calculator can estimate inference using your inputs, but it does not add a fine-tuning training fee. The token counter can help estimate input size before you have production logs. Our words-to-tokens guide explains why a word count is only an approximation.

For an illustrative Gemini 2.5 Flash workload, assume the untuned prompt has 2,000 input tokens, the tuned prompt has 1,000, and both produce 200 output tokens at acceptable quality. No cache, batch discount, extra calls or endpoint fee is included. Using Google’s $0.30 input and $2.50 output rates per million:

Monthly requests	Untuned: 2,000 input + 200 output	Tuned: 1,000 input + 200 output	Tuned plus one $5 training run
10,000	$11	$8	$13
100,000	$110	$80	$85

The first row shows why a cheap training job can still fail to pay for itself at low traffic. In later months with no new training run, the illustrative tuned serving bill would be $8 or $80 respectively. A retraining schedule would add new training charges; real traffic and token counts must replace these invented inputs.

Find the break-even number of requests

In the example, the untuned request costs (2,000 × $0.30 + 200 × $2.50) ÷ 1,000,000 = $0.0011. The tuned request costs (1,000 × $0.30 + 200 × $2.50) ÷ 1,000,000 = $0.0008. The saving is $0.0003 per request. With one $5 training run and no other incremental cost, $5 ÷ $0.0003 = 16,666.67, so the first whole request beyond break-even is the 16,667th. This is arithmetic on a hypothetical workload, not an observed saving or a forecast.

A reusable version is:

Break-even requests = (training + preparation + evaluation + extra hosting) ÷ (baseline cost per request − tuned cost per request)

All terms must cover the same period. If a dedicated endpoint costs a fixed amount each month, add that month’s hosting to the numerator. If you retrain each month, add that charge too. If the tuned version is no cheaper per request, the denominator is zero or negative: there is no positive cost break-even under those assumptions. A tune can still be justified by a measured quality or latency benefit, but the financial case is then a different one.

For example, if this Gemini workload saves only 500 input tokens instead of 1,000, saving per request falls to $0.00015 and the $5 training charge needs 33,334 whole requests to recover. If it needs another 200 output tokens, output cost rises by $0.0005 per request, which more than erases the $0.0003 input saving. Output length can decide the result.

Compare the alternatives first

Try a shorter prompt and a few examples before paying to train. If the same instruction or document prefix repeats across requests, prompt caching may reduce the price of eligible repeated input without a training job. Cache eligibility, writes and discounts depend on the provider; calculate them from actual request patterns.

If the answer must use changing facts from your own documents, retrieval may fit better than a weight update. A tuned model does not automatically receive new documents after training. Retrieval adds its own indexing, storage and lookup costs, so compare the complete workflow. If the job is mainly choosing an existing model, use the AI model tracker to identify a current baseline before building an evaluation set. Then tune only if that baseline misses a measurable target or a verified cost saving survives the whole-bill calculation.

Sources

• Google Cloud Gemini pricing — Gemini 2.5 Flash and Flash-Lite training and text inference rates.
• Google Cloud supervised tuning guide — supported models, training-token calculation and separate evaluation billing.
• Google Cloud deployment guide — tuned Gemini serving and endpoint-charge rules.
• Together AI pricing — Llama 3.1 8B LoRA training rate and job minimum.
• Together AI fine-tuning billing — validation-token calculation and ongoing dedicated-endpoint charges.
• OpenAI API pricing — availability warning for new fine-tuning users.

All pages read on October 29, 2026.

Free toolWhat will 10,000 chats a month cost?Estimate my bill