With rumours of its impending arrival still swirling, Open AI’s new reasoning model has actually arrived.
The first in a new series of “reasoning” models, it is claimed to be closer to human intelligence – and therefore capable of reasoning through complex problems using a multi-step process.
However, its price tag, as predicted, is already attracting some raised eyebrows.
A Much-Hyped Release
The new model is called o1, but The Verge says that it is indeed the much speculated upon Strawberry model.
It is being released as a “preview” and only to ChatGPT Plus and Team users initially. The next group to get access will be the Enterprise and Edu users and that is assumed to be early next week.
This just in! View
the top business tech deals for 2024 👨💻
There is also a companion release – o1-mini, which is a smaller and cheaper option. Access to this model will be made available to all the free users of ChatGPT, but there isn’t a release date yet.
Why Is This Model Different?
The big difference is how this model has been trained. Previous models from the company were taught by mimicking patterns from the training data provided to them.
OpenAI has deployed reinforcement learning for o1, which means that it gets rewarded or penalized for decisions and therefore uses a “chain of thought” process. This makes it a great fit for complex math or coding.
The company reportedly tested its new model out against a team that had qualified for the International Mathematics Olympiad, and it scored 83%. This is compared to the 13% that GPT-4o reached. OpenAI in fact claims that this model performs “similarly to PhD students on challenging benchmark tasks in physics, chemistry and biology.”
The new training method for training also means that o1 delivers less hallucinations, though the OpenAI team did tell The Verge that that they haven’t solved hallucinations entirely.
How Does It Compare to GPT-4o?
The new offering isn’t on a par at the moment with GPT-4o for factual knowledge of the world, says The Verge team. Nor does it have the option to process images or files or browse the web.
However, it does feel “surprisingly human,” says the OpenAI team, and this differentiates it.
But Is the Price Right?
The access to the model is limited at the moment; but it is also not cheap. Developer access costs $15 per one million input tokens, and $60 per one million output tokens. This is considerably more than the costs developers paid for GPT-4o access – which was $5 per 1 million input tokens and $15 per 1 million output tokens.
Maybe with another funding round incoming, OpenAI may hit the jackpot, and the costs can come down for access to Strawberry. Still, it is expensive to create an LLM — and this one is pushing boundaries.
