OpenAI Releases Latest Model OpenAI o1, Confirms Existence of 'Strawberry'
The new model is designed to 'spend more time thinking before they respond', the company said.
OpenAI has released a preview of its latest model, OpenAI o1, finally confirming the existence of the rumoured 'Strawberry'.
o1-preview is the first in a new family of artificial intelligence models that OpenAI is developing, according to the company's press release.
The new model is designed to "spend more time thinking before they respond," similar to human thought processing and rationalisation. It is capable of performing at the same levels as PhD students, on challenging benchmark tests in physics, chemistry and biology, OpenAI said.
For example, GPT-4o can solve 13% of International Mathematics Olympiad problems correctly, while o1-preview can solve 83%.
OpenAI's newest model caters to those working in scientific fields like math, coding, and physics. If true, and if o1 can truly enhance the productivity of such fields, it can prove to be a game changer for researchers across disciplines.
(Source: OpenAI/company website)
In addition to o1, the company is also releasing the o1-mini model. The smaller version is "faster, cheaper reasoning model that is particularly effective at coding". It is also 80% cheaper than the o1-preview model that has currently been released.
Rollout
Starting Thursday night, o1-preview became available for ChatGPT Plus and Team users. The smaller o1-mini model will be available in the ChatGPT model picking interface. There is currently a weekly rate limit, with o1 preview getting 30 messages, while o1 mini gets 50.
It's worth pointing out that the current version that is available to users is just a preview of the full blown model that is currently still in development.
ChatGPT Enterprise and Edu users will see the new model roll out to their accounts in the coming week. While OpenAI intends to make o1-mini free to all ChatGPT Free tier users, like it has with earlier models, it hasn't provided a release date yet.
Hefty Developer Pricing
The advanced reasoning skills from o1-preview come with a hefty price tag for developers. In the API, the o1-preview costs about $15 per 1 million input tokens and $60 per 1 million output tokens, according to The Verge. For context, GPT-4o costs $5 per 1 million input tokens and $15 per million output tokens.
Tokens are chunks of words, characters, or even phrases that are assigned a code while training a model. They're not limited to text alone and can take various data forms. They help AI models understand 'human language', to help formulate the appropriate response.
Notably, incidents of hallucination, the phenomenon where an AI model will pass off false information as true, are much lesser in o1-preview, compared to previous models. But they're not gone completely, "We can't say we solved hallucinations," OpenAI's Research Lead Jerry Tworek told The Verge.
