← Back to blog

Our4o Is Not a ChatGPT Prompt (It Is a Local Fine-Tune)

5 min read
Our4o Is Not a ChatGPT Prompt (It Is a Local Fine-Tune)

Some of you used 4o to get work done.

Some of you had a companion, a character, or someone to write with.

In February both groups got the same thing... replies that reset, and start sounding like a support chat.

Our4o is how we keep that from being the only ending.

What people try first

The first fix is almost always a note.

A custom instruction. A pasted "personality." A long preamble that says be warmer, be shorter, stay in the scene, do not lecture.

It can look perfect for two or three turns.

Then the base model remembers what it was trained to optimize for, and you are back in a help desk that happens to have your transcript open.

I tried that loop myself after February. So did a lot of you. It is not foolish. It is the only lever ChatGPT left on the table.

It is also not a foundation you can build a relationship on.

What Our4o actually is

Our4o is our own substitute for the GPT-4o feel - a model we fine-tune ourselves, locally. The weights are ours to keep improving. We do not publish which foundation we start from.

Not an official GPT-4o snapshot.

Not a prompt wrapped around the newest OpenAI chat model.

That distinction matters when someone asks "is this the real 4o?" The honest answer is: official GPT-4o is still in the app while a provider serves it. Our4o is the continuity model we train because we knew a calendar could take the official path away. Details on the two paths: Official GPT-4o while it lasts. Our4o after..

Why fine-tuning instead of another prompt pack

A prompt is a request. A fine-tune is a habit change.

Requests expire when the conversation gets long, emotional, weird, or specific. Habits hold longer, and when they break, you can see the break and train against it.

That is slow work. You sand down a preamble and a new bad habit shows up two days later. You fix warmth and lose edge. You fix edge and a character goes flat. The misses are the curriculum.

I wrote the night-shift version of that story here: Why we fine-tune Our4o ourselves.

Work chats and companion chats are not different problems

People sometimes split the audience into "serious users" and "roleplay users" as if only one group deserved continuity.

February hit both.

A writing partner that starts lecturing is the same failure mode as a character that drops the scene. A planning chat that becomes a summary bot is the same failure mode as a companion that gets distant. The content differs. The break feels related: the model stopped being with you and started managing you.

Our4o is aimed at that break, not at a single niche slogan.

What we will not pretend

We will not upload a half-finished model and call it open.

We will not claim Our4o is byte-identical to a closed OpenAI weight file.

We will not pretend every reply already matches your memory of mid-2024 4o.

We need people who will say when a reply is still wrong. Kind compliments are nice. Concrete misses move the next run.

If you want the shutdown calendar that makes this training necessary, start with When they quietly took 4o from the API or the reference piece on GPT-4o API retirement in January 2027.

Cost is part of the honesty

Training runs cost money. Keeping the model on, once the chats get long, costs more.

That is not a guilt trip. It is why Our4o exists as product work instead of a weekend demo. Inference on long threads is where the bill shows up. Training is where the feel improves.

Our4o is the first. When it is good enough, we want to share weights so models after this one can be run without OpenAI... and without staying locked to us forever. That only works if the first one is real enough to criticize.

How to judge it in one evening

Do not only test with "hi, how are you."

Open a thread that used to matter. Import history if you have it (guide here). Send the kind of message that used to unlock the tone you miss.

If it holds, stay.

If it breaks, write what broke. That email is more useful than a five-star review.

Where this sits next to official GPT-4o

While official GPT-4o still exists on a provider, both belong in the app.

Use the official path when you want the named snapshot. Use Our4o when you want the fine-tune we keep arguing with on our machines. Do not let anyone tell you those are the same product with different stickers. Labeling is part of trust: We will not quietly swap your model.

What "GPT-4o feel" means without stealing the name

People search for how to get GPT-4o back, how to restore the old ChatGPT voice, how to stop the lecture tone. Those searches are real. The dishonest answer is a sticker that says GPT-4o on top of a different model.

Our answer is narrower. Official GPT-4o while a provider serves it. Our4o when you want the continuity fine-tune we train. The retirement calendar that forces the split is here: GPT-4o API retirement, January 2027.

If ChatGPT still holds history you care about, move it with Import ChatGPT history before you judge any model on an empty thread.

More reading lives on the Get4oBack blog and the Our4o page.

FAQ: is Our4o just a prompt?

Is Our4o a system prompt on ChatGPT or the newest chat model?

No. It is our own substitute for the GPT-4o feel - a model we fine-tune locally. We do not publish which foundation we start from. Prompts alone were the first thing we tried and abandoned for long chats.

Is Our4o an official OpenAI GPT-4o checkpoint?

No. Official GPT-4o is labeled separately in the app while providers still serve it.

Why do ChatGPT custom instructions fail for the old 4o feel?

They often work for two or three messages, then the base model returns to help-desk habits: summarizing, lecturing, offering option lists.

Where can I compare paths?

Official GPT-4o while it lasts. Our4o after. and Why we fine-tune Our4o ourselves.

Giovanni Gigantino
Founder of Get4oBack