ChatGPT or a Custom Model? A Four-Path Decision Tree

Between a monthly ChatGPT subscription and training a model of your own there are four paths, not two, and the gap between the cheapest and the most expensive exceeds eighty times at the same usage volume. The path most business owners picture when they hear "custom model" — training a model on company data — is usually the wrong one, and since May 2026 it is no longer available at all to anyone starting today with the largest provider.

The question is usually posed as "ChatGPT or a custom model?", but the practical answer starts with the nature of the task, not the name of the tool. What follows is the four paths, each with a dated cost, and what you risk in each direction.

The four paths in one table

Start from your situation in the first column and read the whole row. The figures are detailed in the sections below, all in US dollars and current as of September 2026.

The first two rows cover most small and medium businesses in Egypt and the Gulf. The last two need a written justification before you enter them.

One question settles most of the decision: knowledge or behaviour?

The most common confusion is the difference between RAG and fine-tuning, because the two names sound like degrees of the same thing. They are not.

RAG (retrieval-augmented generation) searches your documents at question time, picks the passages that relate to it, and sends them to the model to compose an answer from. The model does not change; what changes is what you place in front of it. The original paper that defined the approach, Lewis and colleagues at NeurIPS 2020, described it as combining memory held in the model's weights with external memory that can be accessed and updated, and concluded that RAG models generate more specific and more factual language than a model relying on its weights alone.

Fine-tuning adjusts the model's weights using examples from your work. The result is a model that writes the way you write, not a model that knows your prices.

The practical difference shows up in a simple question: if a product price changes tomorrow, which path knows the new price? With RAG you update the document and the next request is already correct. With fine-tuning you need a fresh training run, and the old price stays baked into the weights until you finish it.

This is not a matter of preference. A study by Ovadia and colleagues compared the two approaches across knowledge-intensive tasks in different domains and found that RAG consistently outperforms fine-tuning, both for knowledge the model had already seen during training and for entirely new knowledge, and that language models "struggle to learn new factual information through unsupervised fine-tuning".

Specific examples: where each one earns its cost

RAG earns its cost in: an assistant answering store customers from the real catalogue and stock; staff at a distribution company searching pricing and discount policies; a clinic answering questions about doctors' schedules and accepted insurers; any case where you need to know where the answer came from.

Fine-tuning earns its cost in: turning supplier invoices in varied formats into fixed JSON fields that do not break; classifying thousands of messages into predefined categories; replying in a consistent dialect and tone that instructions alone cannot hold, however long you make them.

And there is one classic way to waste money here: a company gathers its internal documents, pays for fine-tuning to teach the model "the company's information", then discovers that answers still invent details and that every document update needs another training run. What was needed was retrieval, not training.

The working rule: always start with RAG. Move to fine-tuning only after running the system proves that what you lack is behaviour, not information.

What changed in 2026: the fine-tuning door is closing at OpenAI

This point alone invalidates half the comparisons published online. OpenAI's pricing page states plainly that the company is winding down the fine-tuning platform and that it is no longer accessible to new users. The official deprecations page sets out the dates:

Inference on fine-tuned models continues until the underlying base model is deprecated, so anyone running a model today does not lose it tomorrow. But for a business owner reading this and planning something new, the meaning is clear: fine-tuning on OpenAI models is not an option on your table at all.

The path remains open at Google. Vertex AI documentation supports supervised fine-tuning on Gemini 3.5 Flash, 3.1 Flash-Lite and the 2.5 family, though not on the newest Flash models. Two clauses in that same documentation deserve the attention of any company considering this path: supervised fine-tuning is not a Covered Service and is excluded from the SLO of any service level agreement, and tuning runs only in us-central1 and europe-west4 while the tuned model is served from US and EU multi-region endpoints only. If data residency is a requirement for you, the decision ends here.

What you actually pay: a company handling 500 customer questions a day

What follows is a hypothetical example with written assumptions for you to replace with your own numbers, not a quotation. The assumptions: 500 questions a day, so 15,000 a month, each consuming 3,000 input tokens (your instructions, the retrieved passages and the conversation history) and 250 output tokens. The monthly total is 45 million input tokens and 3.75 million output, priced at the published rates on the OpenAI, Anthropic and Google pages.

We accessed these prices on 22 September 2026, and they are for standard processing and short context.

Three notes change this table more than the choice of provider does:

The first is that some of these prices are temporary by the provider's own wording. The Gemini 3.8 Flash rate holds until 31 December 2026, then doubles from 1 January 2027, pushing that same row to roughly 95.6 dollars. If you build on a promotional price, put its expiry date in your calendar.

The second is that prompt design lowers the bill more than picking between models does. If a fixed part of your input repeats in every request, such as your system instructions and policies, caching makes it nearly free; Anthropic's pricing page notes that a cache hit costs 10% of the standard input price. And for work that tolerates delay, such as classifying yesterday's messages or processing a batch of invoices, the Batch API gives a 50% discount on input and output alike.

The third is that the embeddings line, which worries many people, is a small one. Indexing a catalogue and policy set of two million tokens at the Gemini Embedding 2 rate of $0.20 per million tokens costs 40 cents once, and embedding a whole month of questions costs under a quarter of a dollar.

The full chatbot bill, including WhatsApp messages charged at Meta's rate by customer country, is broken down in how much an AI chatbot costs.

The economics of fine-tuning run backwards

Training is cheap. The permanent bill is what costs you.

On Gemini 3.5 Flash, the supervised tuning rate is $0.01 per thousand training tokens, and training tokens equal your dataset tokens multiplied by the number of epochs. A set of 2,000 examples at 500 tokens each, over three epochs, gives 3 million training tokens, that is $30 once.

Then the real cost starts. The same page states that a tuned model endpoint is billed at 1.5 times the base model price for the Gemini 3 family onwards. At our example volume, running base Gemini 3.5 Flash costs $101.25 a month and the tuned version about $151.9, a difference of roughly $50.6 a month and $608 a year against a training job you paid $30 for.

Read that the right way: fine-tuning is not rejected because training is expensive, but because it imposes a permanent price tag that a permanent gain has to justify. If the gain is that a smaller, cheaper model now does a job that used to need a larger one, the arithmetic leans your way. If the gain is "answers closer to our style" with no measurement behind it, you are paying $608 a year for an impression.

Where your data goes on each path

What decides your data's fate here is the interface you call and your account tier, more than the provider's name. This is what the official documents say, reviewed on 22 September 2026:

For companies in Egypt and Saudi Arabia there is a practical point that is rarely mentioned: data residency options do not cover your country. OpenAI's region table includes the United States, the European Union, Australia, Canada, Japan, India, Singapore, South Korea, the United Kingdom and the UAE, and neither Egypt nor Saudi Arabia. The UAE is the only regional option, selecting it requires additional approval, and any non-US region requires approval for abuse monitoring controls and a signed Modified Retention amendment, plus a 10% price uplift for models released from 5 March 2026. Anthropic's US-only inference, meanwhile, carries a 1.1x multiplier according to its pricing page.

In practice: the APIs are available in Egypt, Saudi Arabia and the UAE, but processing happens outside your country. If you will be sending customers' personal data, minimize it to what is strictly necessary and strip identifying details before sending, and review your obligations in protecting customer data in Egypt and Saudi Arabia.

When is a self-hosted open model worth it?

Once you run the numbers the answer is clearer than the debates suggest.

An NVIDIA L40S on DigitalOcean GPU Droplets costs $1.57 an hour (prices effective 1 August 2026). Running it for a full month costs $1,130. An H100 reserved for a year costs $3.26 an hour, that is $2,347 a month. And the same page flags a clause many people miss: powering the server off does not stop the billing; only destroying it does.

Compare that with the 500-question table: from $13.50 to $127.50 a month through an API. At this volume self-hosting is nine to over eighty times more expensive. The break-even point against Claude Sonnet 5 sits at roughly nine times the volume, around 4,400 questions a day, and against the cheapest model you never reach it in practice. That is before you count the salary of whoever runs, monitors and updates the servers.

Two genuine justifications remain: data that law or contract forbids from leaving your environment, or volume high and stable enough that a fixed cost beats a meter. Outside those two, self-hosting is a decision you pay for every month in exchange for a feeling of control.

Make the decision once, not again in six months

Start with the narrowest painful task you have, not an assistant that does everything. A single task whose monthly frequency and current cost you know gives you a measure of success before you write a line.

Before launch, prepare a test set of real customer questions with correct answers written by someone who knows the work, and decide the success rate you consider sufficient. Then keep API keys away from any code a user can reach, give the system read permissions limited to the tables it needs, and monitor consumption with a cap and an alert so the bill does not surprise you.

And keep the ability to switch: make the model call a single layer in your code whose provider can change without rebuilding the system. The fine-tuning shutdown dates you read above are exactly why.

If your wider question is the whole system rather than the AI layer alone, the comparison between off-the-shelf and custom at the management-system level is covered in off-the-shelf or custom software. To choose the first task to automate, see which administrative processes to automate first, and for anyone starting from WhatsApp specifically, a WhatsApp chatbot for your store. When you reach the point of hiring a vendor, the questions that reveal how serious they are are collected in how to choose an AI development partner.

Snaabble is a software company in Talkha, Dakahlia, Egypt, working on AI and automation solutions. If you are trying to identify the right path for your case, send your idea to the Snaabble team for an initial plan and an estimate of time and cost within 24 hours, free and with no obligation.

Sources

• OpenAI: Pricing, Deprecations and Data controls in the OpenAI platform
• Anthropic: Pricing and Commercial Terms of Service
• Google: Gemini Developer API pricing, Vertex AI pricing, Gemini API Additional Terms and About supervised fine-tuning for Gemini models
• DigitalOcean: GPU Droplets pricing
• Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, NeurIPS 2020
• Ovadia et al., Fine-Tuning or Retrieval? Comparing Knowledge Injection in LLMs, arXiv 2024