API or your own model: how a startup should build the AI part

An AI product can send its requests to a model that another company runs, or it can run a model of its own. The founder's question is which one the product needs. This page sets out the three options, what each provider's documentation and terms said when read on 26 October 2026, and where Reveneau thinks a startup should start. It is part of the guide on building your AI product after you raise.

The three options

1. Call a hosted model through an API. An API (application programming interface) is the way one program sends requests to another. You send text, the provider's model sends text back, and you pay per request. You shape the answers with written instructions and examples in each request, which is called prompting.
2. Fine-tune a hosted model. Fine-tuning means training an existing model further on your own examples, so that it gives the answers you want with shorter instructions. The provider still runs the model.
3. Run an open-weight model yourself. An open-weight model is one whose trained values, called weights, are published so that anyone can download and run it under a licence. You run it on servers you rent or own.

OpenAI's own guide to improving model results starts with the first option. It says: "The prompt engineering process may be all you need in order to get great results for your use case" [1]. Prompt engineering is the work of writing and testing those instructions.

Option	What you control	What the sources say on 26 October 2026
Hosted model through an API	Instructions, examples, which model	All three large providers sell this; models are retired on the provider's schedule
Fine-tuned hosted model	Training examples	Closed to new users at OpenAI; no tunable model in the Gemini API; offered on Google Cloud
Open-weight model on your servers	Everything, including where data goes	gpt-oss-120b fits on a single 80GB GPU by its model card; licences differ by model

Fine-tuning at OpenAI and Google

OpenAI's pricing page says: "OpenAI is winding down the fine-tuning platform. The platform is no longer accessible to new users, but existing users of the fine-tuning platform will be able to create training jobs for the coming months" [2]. Its deprecations page gives the dates [3]:

Date	What OpenAI's deprecations page says
May 7, 2026	Fine-tuning is no longer available to organizations that have not run it before
July 2, 2026	Fine-tuning is no longer available to organizations that have not used a fine-tuned model in the past 60 days
January 6, 2027	Active existing customers can no longer create new fine-tuning jobs

Models already tuned keep working for a time: "Inference on fine-tuned models will continue to be available until the base models are deprecated" [3]. Inference means using a model to produce answers, and a deprecated model is one the provider has scheduled for retirement. For existing users, OpenAI still lists prices, for example $5.00 per million tokens for training gpt-4.1-mini, with $0.80 for input and $3.20 for output when the tuned model is used [2]. It adds that "Inference discounts are available if you enable data sharing when creating the fine-tune job" [2], which means OpenAI receives your data in exchange for a lower price.

Google closed fine-tuning in its Gemini API earlier. Its documentation says that "With the deprecation of Gemini 1.5 Flash-001 in May 2025, we no longer have a model available which supports fine-tuning in the Gemini API or AI Studio", and points to its enterprise platform on Google Cloud [4]. It adds: "We don't have immediate plans to bring fine-tuning support" back to the Gemini API [4].

On Google Cloud, supervised tuning was listed on 26 October 2026 for Gemini 3.5 Flash-Lite, Gemini 3.5 Flash, Gemini 3.1 Flash-Lite, Gemini 2.5 Pro, Gemini 2.5 Flash-Lite and Gemini 2.5 Flash [5]. Supervised tuning means training on pairs of example inputs and the outputs you want. Google's guide says: "For best results, provide at least 100 to 500 examples" [5]. Google's Gemini API terms say that "Google only uses content that you import or upload to our model tuning feature for that express purpose" [12].

Anthropic's developer documentation index, read on 26 October 2026, lists no page about fine-tuning: a text search of the index for the word finds none [6]. That shows only what Anthropic's own documentation lists. This guide did not check what Anthropic models offer on other clouds. Anthropic's pricing page gives its own advice on choosing among its models: "Choose Haiku for simple tasks, Sonnet for most production workloads, and Opus for the most complex reasoning" [7].

Open-weight models and their licences

An open-weight model can run on servers the company controls, so no request leaves for a provider. The licence decides what you may do with it.

OpenAI's model card for gpt-oss-120b lists its licence as Apache 2.0, an open-source licence [8]. The card describes gpt-oss-120b as made for production and general use, with 117 billion parameters of which 5.1 billion are active, and says it fits on a single GPU with 80GB of memory, naming the NVIDIA H100 and AMD MI300X [8]. Parameters are the numbers inside a model that training sets, and the active parameters are the part used for each piece of text. A GPU is the type of chip that runs AI models. The smaller gpt-oss-20b has 21 billion parameters with 3.6 billion active, and the card describes it for lower latency, meaning faster answers, and local or specialised uses [8]. The card also warns that both models "should only be used with the harmony format as it will not work correctly otherwise" [8]. The harmony format is the way requests and answers must be laid out for these models.

Meta's Llama 3.1 Community License adds a condition for large products. If a product had more than 700 million monthly active users in the preceding calendar month, the licence says "you must request a license from Meta, which Meta may grant to you in its sole discretion" [9]. The licence has other terms as well, so read it whole. This guide read the Llama 3.1 licence only, and newer Llama versions may differ.

Google's Gemma Terms of Use, last modified April 1, 2026, send users of Gemma 4 to a separate page: "For Gemma 4 terms, see the Gemma 4 license" [10]. The address of the linked page contains "apache_2". This guide did not open that page.

Running your own model has its own costs: the servers, the people who run them, and the work of updating the model. This guide found no published server price to compute it from, so it gives no dollar figure. For a startup that needs this, Reveneau's service page on your own AI model server describes that build, and the open-source directory lists open projects by the paid tool they replace.

The provider's terms limit what you build

Read the provider's terms for limits on your product, as well as on your data. Anthropic's commercial terms say a customer may not "access the Services to build a competing product or service, including to train competing AI models or resell the Services except as expressly approved by Anthropic" [11]. Those terms took effect on June 17, 2025. A startup that plans to use one model's answers to train its own model should read each provider's terms on that point before it starts.

The page choosing a model provider: terms and data use compares what each provider keeps and for how long.

Hosted models are retired on a schedule

A hosted model can be retired, and a product built on it has to move. OpenAI's deprecations page, for example, says "The gpt-5.4-cyber model is deprecated and will be removed from the API on October 1, 2026" [3]. Anthropic's pricing page marks some older models as "Retired" and adds that they "May still be available on other cloud platforms" [7].

The same OpenAI page says its Evals dashboard and API "are scheduled to shut down" on November 30, 2026 [3]. Evals are written, repeatable tests of a model's answers. Keep your prompts and your eval cases in your own files, in your own code repository, so that a retired model or a closed tool means running the same tests against a new model. The guide what happens when the AI model changes sets out what to rerun, and investors look at this risk too, as the page on model dependency risk explains.

Reveneau's view: where a startup should start

Reveneau's view is that a startup should build its first version on a hosted model called through the API. Put the work into three things: clear written instructions with examples in each request, a choice of model sized to the task, and an eval suite that measures the answers on your own cases. Move to one of the other options only for a reason you can name:

• Fine-tune when the eval suite shows a gap that better instructions have failed to close, and the provider still offers tuning for the model you need. Collect the examples first. Google's guide recommends at least 100 to 500 for best results [5].
• Run an open-weight model when a customer contract requires that data stays on servers you control, or when your own measured costs at your own volume favour it. Check the licence of the exact model and version.
• Change hosted models when the evals show another model passes more of your cases, or when your current model is retired.

Each move is a decision you can test: run the same eval suite before and after, and compare. The guides on why AI evals matter and LLM evals explain how to build that suite, and the page what an AI feature costs to run covers the cost side.

Where Reveneau fits

Reveneau is an AI software development consultancy with offices in San Francisco and Vancouver. Reveneau says AI writes all of its code and every change must pass an eval suite written from the specification. It says Jev, TypeSafe AI's decision model, grades that suite, and that the suite runs ten times faster than it did with its previous grader. Reveneau says it uses AI instead of hiring more engineers and puts that saving into clients' prices, and that it takes responsibility for the whole project through production and after release. To plan which model your product should use, see AI development at Reveneau or contact Reveneau.