OpenAI is about to drop o3-mini. Here's what to expect.

• OpenAI's o3-mini will launch in roughly two weeks.

• It will be available via API and ChatGPT with 'real high' limits

• It's expected to be better value than o1-mini.

Last month, OpenAI announced it had created a high-performing reasoning model with benchmarking scores that approach human levels of cognitive ability.

By the company's own account, the "o3" family will be a major leap towards the goal of artificial general intelligence, which would give a model the ability to learn new skills and solve complex problems.

Like previous model families, OpenAI will release a full version and a mini version (and possibly a "Pro" version). It's not yet clear when the full model will be launched, but CEO Sam Altman says o3-mini should drop in the next couple of weeks.

There's no official performance data available for o3-mini, but Altman says it's "FAST," but worse than o1-pro at most things. o1-pro is the most advanced model in OpenAI's o1 group and is only available via a $200-a-month subscription.

Immediate API access and 'real high' limits

Performance aside, the launch comes with plenty of bonuses for indie hackers.

Unlike previous releases, OpenAI will launch o3-mini across both its API and ChatGPT interface at the same time, giving developers the chance to play with it right away.

It's not clear how much API usage will cost, but Altman hinted the ChatGPT version will have greater usage limits. He didn't put a number on it, but he did tell X users the o3-mini message rate would be "real high."

ChatGPT Plus and Teams users currently get up to 50 messages per day with o3-mini's predecessor, o1-mini.

We don't yet know exactly who will be able to use o3-mini, but Altman has confirmed Plus users will get access. Pro access is therefore pretty much a given. But Teams and Enterprise subscribers will have to wait and see.

Reasoning models

Despite the name, o3 is the second generation of OpenAI's "reasoning models," which are designed to think more like humans. They're given more time to "think" about an answer, and can solve more complex problems and correct their own mistakes as they produce a final response.

o1 performs far better than the company's non-reasoning models on benchmarking tests for things like scientific reasoning. Although it started out sluggishly, with a preview version taking up to a minute to answer some complex questions, a full version released in December was much, much faster.

The reasoning model has prompted mixed responses from some users — particularly those on "Pro" — sometimes taking a long time to produce results no better (or even worse) than OpenAI's older GPT models.

But, as Dawn Analytics co-founder Ben Hylak explains at length in a post on Latent Space, o1 simply isn't a chat model and shouldn't be used in the same way. To get the best results, he recommends providing far more detailed prompts and being clear about the goals of your requests.

Altman says o3-mini almost always offers a lower performance than o1-pro. That implies it's maybe comparable on certain benchmarks. A promising indicator for the full o3.

In fact, OpenAI is so confident in the upcoming model that it's shifting its wider goals from AGI to ASI — artificial superintelligence. This is the kind of thing you see in sci-fi movies: AI that significantly outsmarts people. Altman says it will unlock incredible scientific progress and "massively increase abundance and prosperity."

It's exciting (and scary), but probably won't have that much direct impact on indie hackers any time soon — besides ramping up progress on smarter, cheaper models for all.

For indie hackers, it's likely the next most exciting step for OpenAI will be the release of it's long-anticipated AI agents, which should be primed to automatically perform more focused tasks. Although the company hasn't explicitly mentioned their progress since a scoop from Bloomberg last Fall, Altman recently said he expected agents to "join the workforce" and "materially" change company outputs.