What and why

You may be tempted to purchase(or are unable to afford) the $20/month subscription for ChatGPT plus. However, depending on the range of your usage and frequency of usage, it may not be so good of a deal compared to what you can just get from the API, which is pay-per-use.

The OpenAI API(OpenAI Platform) is intended for usage of developers, who may want to add the term "AI" to their products with no actual improvements to user experience but makes brand sound better somehow. With that being said, you do not need to know how to code to use most of the OpenAI API - the developer platform fortunately provides a playground where you can use it as if it is ChatGPT, with some additional settings you can tweak.

Pricing

Instead of paying a set $20/month, how much it costs depends on your usage. The OpenAI api bills you based on the amount of tokens you use. Tokens are how LLM AIs see words - in English they are usually made up of several letters but can take up several parts of a word. You can use the tokenizer tool from OpenAI to understand the concept, though in practice it won't matter too much.

Usually, basic models such as gpt4o/gpt3/gpt4 will likely cost you less than a cent per run. Heavier models such as o1 or gpt5 are likely to cost you about 10 cents per run, due to having reasoning(even more so if you turn on reasoning summaries). This is my estimate for the models' response to a simple question (e.g. "what is xxx") - asking it for more complex tasks may, obviously, cost more.

Although the final response is shorter, GPT5 tends to think really hard about its respsonse and hence incurs more usage. You can tune down the reasoning if it's unnecessary. This run would cost me 0.8 cents.

Limits

It may seem like that if you intend to mainly use the heavier models, $20/month may be a better option for you, assuming your API usage will go that much. However, even paid accounts can have limits on how many times a model can be used per 3 hours[1]. This seems like more than enough, but then again, if you are not using that much, why pay for that much?

For GPT5 specifically, they implement a model router that will automatically set the models reasoning effort and verbosity levels based on your request[2], and it's been reported to be not so accurate[3]. Over the API though you can explicitly ask the model to think as hard as it can.

Using the API however is still subject to their rate limits and how much you have in your wallet. The more money you have spent in total in the API, the higher the limits will be(by getting to a higher tier), even if you are unlikely to hit those limits in normal usage. Did I mention their rewards program where if you give them your data, you can get $2.5 worth of free tokens a day for tiers 1-2, and $10 worth of free tokens a day for tiers 3-5, for bigger models like gpt5 or 4o?

The fun part

Using the API doesn't mean you need to stop using your free account - in fact you should be using that first. Assuming you are a student that needs to upload pictures to ChatGPT 10-15 times a day, always use your free daily uses on ChatGPT, then switch over to the API for any additional uses. You are then on a pay per use basis for any usage outside of your free access.

Caveats

Not everything available in ChatGPT can translate 1:1 over to the API. Below is a table of what's available(and not) and possible replacements.

ChatGPT Features	OpenAI API
Memory	❓Kinda?*
Custom Instructions	❓Kinda?*
Web Browsing	✅Yes
GPTs	❓Kinda?
Reading image and audio	✅Yes
Reading other files	✅Yes**
Voice	✅Yes
Chat History	❓Kinda?
Advanced Voice(video etc.)	❌No
Image Generation	✅Yes
Sora	❌No
Deep Research	✅Yes

*You can preset them in system instructions. Or upload them in a file and ask it to explicitly go over it if you would like it to understand your entire life up to their point.

Math Equations

The AI agents will write math equations in LaTeX, which the platform does not render. There are several methods for this:

ChatGPT Method

The easiest way to do this is to paste the output directly into normal ChatGPT(chatgpt.com) and ask it to "repeat the following as is". Now you have equations that are rendered properly.

Markdown Method

Paste the output into a markdown editor(such as Obsidian) and modify the equations so that they can be rendered. For example:

A linear transformation is a function between two vector spaces that preserves two operations:
- Addition: ( T(\mathbf{u} + \mathbf{v}) = T(\mathbf{u}) + T(\mathbf{v}) )
- Scalar multiplication: ( T(c\mathbf{v}) = cT(\mathbf{v}) ) for any vectors ( \mathbf{u}, \mathbf{v} ) and any scalar ( c ).

For inline equations like these, you will have to figure out where they start and end, and append a $ to both ends of the equation, like so:

- Addition: ( \$\T(\mathbf{u} + \mathbf{v}) = T(\mathbf{u}) + T(\mathbf{v})\$ )

You may also encounter full-line equations as follows:

[
T\left(\begin{pmatrix} x \ y \end{pmatrix}\right) = \begin{pmatrix} -y \ x \end{pmatrix}
]

Simply swap the brackets out for $$.

\$\$
T\left(\begin{pmatrix} x \ y \end{pmatrix}\right) = \begin{pmatrix} -y \ x \end{pmatrix}
\$\$

Once you make the change and paste it into an editor, you should be able to see the math equations.

Instructions Method

Alternatively, give it the following prompt(preferably in system instructions so it would listen):

Features

Web Browsing

This is only available in the prompts tab, not for assistants.

You will need to manually enable this feature. Go to tools - create and add web search in settings then fill it out(or not).

Reading non-image/audio files

Although images or audio can be used as if it were part of the text input, reading documents is an external feature just like web browsing. Go to tools - file search then drag your document in. Then just ask it like normal(you may need to explicitly ask it to search the file, since it is treated as a vector store).

As of writing you can now attach PDF files directly alongside images or audio(the rest still needs to go to file search). Although note that openai processes PDF files by attaching both the file in its text form and screenshots of the file - where I personally just attach only the latter.

GPTs

You can't access GPTs that others have created for ChatGPT. You can however use assistants to create a GPTs-like experience, where you can create multiple presets(simlar to prompts) of agents, each with different settings and system messages. The platform also provides a way to better phrase your responses with AI in the editor for free !! it costs money now ☹.

Code Interpreter

ChatGPT can sometimes run code in a sandboxed environment to analyse stuff or generate stuff. You can do the same in the API via Assistants and enabling it in tools.

This is a bit anticlimatic, but I was not able to download the file myself. Perhaps the API docs has more information on how to get it. Also note that using tools like this adds steps to the AI internally(the agent has to use the tool then read from the tool to produce results) hence it costs a lot more tokens.

Chat History

The OpenAI playground does not store your conversation history like ChatGPT would. Your conversations are stored only on the device you made the request on, is not designed to persist(can disappear after clearing cookies for example), and is not stored if your request has any files attached. You can still enable storing logs when you make the request(and you can see those logs), but it's a bit of a hassle to paste it back into the prompt editor to recreate your conversation as is.

Image Generation

You can add gpt-image-1 to your agent as a capability(Image Generation). From testing it still can't generate a full wine glass.

Deep Research

Select either the o3 or o4-mini deep research model from the model list. For this one you will have to add the web search tool and set the version to deprecated.

When you use ChatGPT's deep research, it will always first use a standard chat model to confirm the direction of your research by asking you a few questions, before switching to the actual deep research model. Although you can replicate this behaviour yourself by switching models halfway, the prompt for the first agent is not available(because I don't want to get sued by OpenAI). I really don't think it's too hard to spin up a prompt for it though.

Warning

Make sure you disable any of their pro models in the organisation settings!! These models usually cost $5 per run which was not worth my money to find out.

References

[1] https://help.openai.com/en/articles/11909943-gpt-5-in-chatgpt
[2] https://openai.com/index/introducing-gpt-5/#:~:text=One%20unified%20system,handles%20remaining%20queries
[3] https://www.reddit.com/r/singularity/comments/1mkreom/the_gpt5_router_is_a_bad_joke/
[4] https://platform.openai.com/docs/guides/pdf-files?api-mode=responses#usage-considerations