What is fine-tuning in plain terms?

Fine-tuning is an additional training step that starts with a pretrained model and adds task- or domain-specific information[3]. It can change how the model responds to a type of request. NIST describes generative AI training as often including foundation-model pretraining and fine-tuning, with supervised fine-tuning or human-feedback methods used after initial training[1]. It is a model adaptation, not a search through the latest uploaded files.

Can a new OpenAI customer start fine-tuning in October 2026?

OpenAI says its fine-tuning platform is winding down and is no longer available to new users[4]. Eligible existing customers can still create training jobs for a limited period, subject to the base model's availability. A business planning a new project should verify eligibility before budgeting for training. File search is a separate retrieval tool and may fit a task where the relevant knowledge lives in documents that need updating.[5]

Which GPT-4.1 snapshots appear in the fine-tuning guide?

OpenAI's supervised fine-tuning guide lists gpt-4.1-2025-04-14, gpt-4.1-mini-2025-04-14 and gpt-4.1-nano-2025-04-14 [4]. A name in the guide is not a promise of indefinite access. An existing customer should check the platform notice and its own account before starting another job, then record the base model used by each fine-tuned version. Its retirement date affects whether the resulting model can continue serving requests.

How does Responses API file search find an answer?

File search retrieves relevant information from uploaded files using semantic and keyword search over a vector-store knowledge base[5]. The application supplies a vector store containing files, and the Responses API can call the file-search tool while generating a response. This leaves the documents in a retrieval system rather than adapting the base model through another training job. The business should still check that retrieved passages support the answer it gives a user.

What does OpenAI file search cost?

The cited OpenAI price page lists storage at $0.10 per GB per day after 1 GB free, and tool calls at $2.50 per 1,000 calls[6]. Tokens consumed by the tool and model follow the chosen model's token rates. A budget therefore needs three meters: stored volume over time, number of tool calls and model-token usage. The published rates are inputs to a calculation, not a fixed monthly bill.

Does file search replace every fine-tuning project?

No. Fine-tuning changes a pretrained model through task-specific training; file search retrieves information from uploaded documents[3][5]. If the desired change is a stable response pattern, the training question differs from keeping a changing document collection available to an answer. The buyer should test the actual task, outputs and update frequency. OpenAI's wind-down also changes whether a new fine-tuning project is available to that customer at all.[4]

What Is Fine-Tuning? Availability and File Search Costs

Quick Answer: Fine-tuning further trains a pretrained model; file search retrieves documents.³^,5 OpenAI bars new users; eligible existing users can train until 6 January 2027 unless the base model closes earlier.⁴ File search costs $0.10 per GB-day after 1 GB free, $2.50 per 1,000 calls, plus model tokens.⁶

Summary in a mind map

What Is Fine-Tuning? Availability and File Search Costs
│
├─ Model adaptation
│   ├─ Start with a pretrained model
│   ├─ Add task-specific training
│   └─ Test the resulting behaviour
│
├─ Platform access
│   ├─ Fine-tuning is winding down
│   ├─ Existing access is time limited
│   └─ Name the exact GPT-4.1 snapshot
│
├─ Document retrieval
│   ├─ Upload files to a vector store
│   ├─ Search by meaning and keywords
│   └─ Review answers against passages
│
└─ Budget the choice
    ├─ Storage after the free allowance
    ├─ Tool calls per task
    └─ Model tokens at chosen rates

What fine-tuning changes

The question “what is fine-tuning?” concerns a training stage, not a document lookup. NIST's CSRC glossary defines it as starting with a pretrained model and adding task- or domain-specific information³. NIST's 2025 report explains that generative AI training often has a foundation-model pretraining stage and a later fine-tuning stage. It also describes supervised fine-tuning and reinforcement learning from human feedback after initial pretraining as ways to align the base model with human preferences¹.

A dental practice might want a model to follow a consistent classification pattern for incoming administrative messages. That is an illustrative task, not evidence that its data are suitable or that fine-tuning will improve the outcome. The team would need example inputs, expected outputs and a held-out test. The point is to define the desired behaviour before paying for training. If the problem is simply finding the current instructions in a changing set of policy files, retrieval may ask a more direct question.

Fine-tuning is sometimes used as a catch-all label for any AI customisation. That blurs an important budget choice. A model trained to respond in a pattern does not automatically contain the latest version of a practice's documents. File search lets an application look up uploaded material at response time⁵. The two paths should therefore be compared on the actual job: stable behaviour, changing knowledge, or both.

OpenAI wind-down in October 2026

OpenAI says it is winding down its fine-tuning platform and has closed access to new users⁴. Eligible existing customers can create training jobs for a limited period. Its current notice gives a January 6, 2027 end point for those jobs unless the base model closes earlier⁴. This is a platform availability statement, not a general claim that every fine-tuned model stops serving on the same day. The business should confirm its own eligibility and base-model schedule before committing to a project.

That timing matters most to a new buyer. A proposal built around an unavailable training feature is not executable even if the technical idea is sound. An existing customer has a different question: whether the intended model remains available long enough to train, test and use the result. The project record should name the account, base model, intended use and the check date. Do not infer permanent access from a completed training job.

For a related view of where an autonomous system may be worthwhile, see AI agents for small businesses. The fine-tuning wind-down changes one model-training option. It does not decide the broader question of whether a workflow should use an agent, an automation platform or a simpler search interface.

GPT-4.1 snapshots in the guide

OpenAI's supervised fine-tuning guide names three GPT-4.1 snapshots: gpt-4.1-2025-04-14, gpt-4.1-mini-2025-04-14 and gpt-4.1-nano-2025-04-14 ⁴. The exact suffix matters. A project plan saying only “GPT-4.1” may hide which base model the training job and resulting model use. The business should record the exact name and check the platform's availability notice before scheduling work.

Snapshot named by OpenAI	Buyer check
gpt-4.1-2025-04-14	Base and fine-tuned versions shut down 14 January 2027 ⁴
gpt-4.1-mini-2025-04-14	Base and fine-tuned versions shut down 14 January 2027 ⁴
gpt-4.1-nano-2025-04-14	Base and fine-tuned versions shut down 23 October 2026 ⁴

File search retrieves documents

The Responses API file-search tool looks for information in uploaded files using semantic and keyword search⁵. OpenAI describes a vector-store knowledge base for that retrieval. The application can attach the tool to a response request; the model can use retrieved material to answer. This process differs from teaching a base model a new response pattern through a training job. Updating a document collection is a content-management problem, not a reason by itself to retrain the model.

For the practice, a test might use an approved reception procedure and ask a narrow question whose answer appears in that procedure. The reviewer should inspect the cited passage, the answer and any omitted exception. A retrieved paragraph can be relevant yet insufficient. The team should test that the answer does not generalise beyond the document or present an out-of-date version as current. File search supplies retrieval; the practice still owns the correctness of its own documents and the response it sends.

Set up the vector store

OpenAI's file-search guide says a vector store containing uploaded files must be set up before the tool is used with the Responses API⁵. The business needs to decide which documents belong there and who can replace them. A version label and approval date are useful operational details even though the cited guide does not prescribe the practice's internal review process. They allow a reviewer to identify which document the tool could have found when a particular answer was generated.

The first trial should contain a small, known set of files. Record the question, expected source passage, returned answer and reviewer judgement. Then add a revised document and repeat. This is a practical way to detect whether an outdated answer comes from a stale file, a retrieval miss or the generation step. The vector store is part of the system's knowledge base, but it does not make every answer reliable without testing.

Calculate file-search cost

OpenAI lists file-search storage at $0.10 per GB per day after 1 GB free, and Responses API file-search tool calls at $2.50 per 1,000 calls. Tokens used by the tool are billed at the chosen model's token rates⁶. The budget therefore has three components. Estimate stored volume over the period, count tool calls and measure model-token use in representative requests. These are rates as checked in October 2026, not a guarantee that future prices will stay unchanged.

Cost component	Published rate or meter	What the practice records
Stored files	$0.10 per GB-day after 1 GB free	Average stored volume and days
File-search calls	$2.50 per 1,000 calls	Calls per completed task
Model tokens	Chosen model's rates	Input and output tokens per task

For an example budget, suppose a pilot stays within the free storage allowance and makes 1,000 file-search calls. The tool-call line would be $2.50 at the cited rate; model tokens remain additional⁶. That arithmetic is illustrative. The practice should measure its own call pattern because one user question may trigger a different number of calls from another. A monthly bill is not the storage rate multiplied by file size alone.

What this does not decide

The cited sources do not show which method will answer this practice's messages accurately, whether its existing OpenAI account is eligible for further training, or the price of its chosen model's tokens⁴^,5^,6. A document retrieval test and a model training test require different evidence. Compare them using the same task and a reviewer who knows the correct answer. The announced wind-down is a practical constraint on availability; it is not evidence that every retrieval design is superior.

For either trial, the practice should control who can change its test files and examples. NIST SP 800-218A recommends protecting training, testing, fine-tuning and alignment data against unauthorised access and modification, while monitoring the confidentiality and integrity of non-public data².

What to do next

Check the existing account's fine-tuning access and the exact snapshot before planning another job. If the task depends on current files, create a small vector store, test retrieval against approved documents and record all three cost meters. For deciding which business workflow deserves the effort, the agent opportunity score and accounts payable automation example provide adjacent comparisons.