The Fine-Tune You Keep Paying Rent On
Fine-tuning is sold as a training cost. On every platform that still offers it, the training is the small number and the hosting is the bill.
This note reads the published fine-tuning and customisation pricing at OpenAI, Google and AWS, as the pricing pages showed it on October 30, 2026.

Two of the three large developer APIs have stepped back from the feature entirely. The one route that prints a monthly figure prints a large one.
Two developer APIs are no longer taking new fine-tuning customers
The OpenAI pricing page says it above the table rather than hiding it. OpenAI is winding down the fine-tuning platform, and the platform is no longer accessible to new users.

Existing users will be able to create training jobs for the coming months, per the same notice. All fine-tuned models will remain available for inference until their base models are deprecated.
Google's developer API has been in that position for longer. Its fine-tuning page says that with the deprecation of Gemini 1.5 Flash-001 in May 2025, no model available in the Gemini API or AI Studio supports fine-tuning.
It adds that there are no immediate plans to bring fine-tuning support back, while inviting feedback from teams for whom it matters. Both statements are on the page checked here, not inferred.
So a team starting today cannot fine-tune on either of those two surfaces. The remaining routes are the enterprise clouds, which price the work differently.
What is left on the OpenAI table is one model and an hourly rate
The fine-tuning section lists a single model row: o4-mini-2025-04-16. Training is priced at $100.00 per hour, and inference on the result at $4.00 input, $16.00 output and $1.00 cached input per million tokens.
An hourly training rate is the awkward part of that line. Dataset size does not translate into a bill, so the training cost is not knowable in advance from anything you control.
Inference has two discounts and they compound. With data sharing enabled the rates become $2.00, $8.00 and $0.50, which is exactly half, and the batch table halves the standard rates to the same figures.
- Standard: $4.00 input, $16.00 output, $1.00 cached input
- Standard with data sharing: $2.00, $8.00, $0.50
- Batch: $2.00, $8.00, $0.50
- Batch with data sharing: $1.00, $4.00, $0.25
Training stays at $100.00 per hour in all four rows. The page notes that inference discounts are available if you enable data sharing when creating the fine-tune job, so the choice is made once, at the start.
One comparison on the same page is worth making. Fine-tuned o4-mini inference at $4.00 and $16.00 sits above gpt-6.1-terra at $1.00 and $5.00 standard, a current flagship row on the same rate card.
That is editorial arithmetic on two published rows rather than a performance claim. It does mean the cheaper path is not automatically the customised one.
Vertex prices training by the token, which can be forecast
The Vertex AI table uses training tokens instead of training hours, and defines the unit. Training tokens are the total number of tokens in your training dataset multiplied by your number of epochs.
The rates in the per-million-token tables are small. Gemini 2.0 Flash shows tuning training at $0.003 per million tokens, and Gemini 2.0 Flash Lite at $0.001.
Older rows are priced per thousand instead, and much higher. Gemini 1.5 Flash shows a training token price of $0.008 per 1,000 and Gemini 1.5 Pro $0.08 per 1,000.
The serving side is the good news on this platform. The footnote says a tuned model endpoint has the same prediction price as the base model, so there is no standing hosting premium in a global region.
That exemption has a geography attached. A separate footnote says a tuned model endpoint will be 1.5 times the base model in a non-global region, so the serving premium depends on where the endpoint sits.
Bedrock is where the monthly rent appears in print
The Amazon Bedrock page publishes customisation as three columns rather than one, and the column names are the whole story.
For Meta models they read: price to train 1M tokens, price to store each custom model per month, and price to infer from a custom model for one model unit per hour with no-commit provisioned throughput.
- Llama 2 Pretrained (13B): $1.49 to train 1M tokens, $1.95 per month to store, $23.50 per hour to serve
- Llama 2 Pretrained (70B): $7.99 to train 1M tokens, $1.95 per month to store, $23.50 per hour to serve
The storage figure is a footnote-sized number. Custom model storage is $1.95 a month regardless of model size, which is the cheapest line in the whole exercise.
AWS supplies its own worked example, and it is more useful than the table. For a 70B tune on 1,000 tokens of data, it totals training, one month of storage and one hour of inference at $33.44.
Then the same example turns to hosting the result properly. Provisioned throughput on a one-month commitment is $21.18 per hour per model unit, which AWS multiplies out to $15,757.92 per month.
The Cohere example on the same page is larger again. It totals $55.45 for training, storage and an hour of evaluation, then prints $29,462.40 a month for one model unit at $39.60 an hour.
Both of those totals are the vendor's own arithmetic, printed on the pricing page. They are the clearest statement anywhere on these pages of what a custom model actually costs to keep.
On one route the custom model has no per-token option at all
This is the constraint that turns hosting from a choice into a requirement, and it is a single asterisked line under the Meta tables.
Llama 2 pre-trained models are available only in provisioned throughput after customisation, per that note. There is no pay-per-token row for the customised model to fall back on.
Amazon's own models work the other way. The page says on-demand inference for custom Nova models is priced the same as base Nova inference, which removes the monthly commitment entirely.
The same page lists fine-tuning tabs for Llama 3.1 and Llama 3.2, but their customisation figures did not render in the page markup on the date checked. The Llama 2 table was the only one readable.
A custom model inherits the shutdown date of its base
The OpenAI notice is explicit that a fine-tune does not extend a model's life. Fine-tuned models remain available for inference until their base models are deprecated.
That makes a customisation project a dated asset rather than a permanent one. The retirement schedules it inherits are in the model shutdown calendar.
It is also why the hourly hosting figures matter more than the training ones. A commitment priced by the month has to be re-earned against a base model that will eventually be withdrawn.
What to price before committing to a fine-tune
- Whether the platform still accepts new fine-tuning customers, because two of the three do not
- Whether training is billed by the hour or by the token, since only one of those can be forecast
- Whether the tuned endpoint carries a serving premium, which is zero on one platform and 1.5 times on the same platform outside the global region
- Whether a per-token option exists after customisation, or only a monthly commitment
- The storage charge, which is trivial, against the hosting charge, which is not
- The deprecation date of the base model, because the custom model shares it
The commercial pattern here is familiar. A low entry figure sits next to a standing charge that only appears further down the page, which is the subject of the add-on note.
Bottom line
Fine-tuning is now an enterprise cloud feature. The two large developer APIs that small teams actually start on have either closed the door to new users or removed the capability.
Where it remains, training is cheap and serving is not. One vendor's own example puts a single model unit at $15,757.92 a month, against $1.95 to store the weights it serves.
For most small teams the honest comparison is against a good prompt on a current flagship, which costs nothing to keep and retires on someone else's schedule.
Assistant plans that avoid this decision entirely are graded in the Cursor review and the GitHub Copilot review, with the shortlist in the assistant plan guide.
For what happens when a vendor wants a conversation before quoting any of this, see the trial that needs a sales call and the monthly cap note.
Sources
- OpenAI API pricing (fine-tuning section and wind-down notice) - checked October 30, 2026
- Gemini API fine-tuning documentation - checked October 30, 2026
- Vertex AI generative AI pricing (tuning training and tuned endpoints) - checked October 30, 2026
- Amazon Bedrock pricing (model customisation and provisioned throughput) - checked October 30, 2026