Openai on Friday launched a novel model of “reasoning” AI, O3-Mini, the latest in the family of Models of the reasoning of the O.
Opeli for the first time announced the model in December along with a more talented system called O3, but the premiere takes place at the key moment of the company whose ambitions – and challenges – apparently grow overnight.
Opeli is fighting the belief that this is the basis in the AI race with Chinese companies, such as Deepseek, which, according to OpenAi, could steal his IP address. Tried Keep your relationship with Washington Because at the same time he implements an ambitious project of the data center, and, as it is supposed to be the basis for one of the largest funds of financing in history.
Which leads us to O3-Mini. Opeli raises its novel model both as “powerful” and “affordable”.
“Today’s starting signs […] An important step towards expanding the availability of advanced artificial intelligence in the service of our mission, “said Opeli spokesman, said Techcrunch.
More proficient reasoning
Unlike most huge language models, the reasoning of models such as O3-Mini thoroughly check the facts before giving the results. This helps them avoid some traps that usually stumble on models. These reasoning models last a little longer to reach solutions, but the compromise is that they are more reliable-although not perfect-in fields such as physics.
O3-Mini is adapted to STEM problems, especially in the field of programming, mathematics and exact sciences. Opeli claims that the model is largely equally with the family of O1, O1 and O1-Mini, in terms of possibilities, but it works faster and costs less.
The company claimed that external testers preferred O3-Mini answers from the answer from o1-mini over half of the time. O3-Mini apparently made 39% less “main mistakes” on “difficult questions” A/B tests. Versus O1-Mini and made “clearer” answers, while at the same time providing a response of about 24%.
O3-Mini will be available to all users via CHATGPT from Friday, but users who pay for Chatgpt Plus and the OPENAI team plans will obtain a higher rate limit of 150 queries per day. CHATGPT PRO subscribers will receive unlimited access, and O3-Mini will come to ChatgPT Enterprise and ChatgPT EDA for a week. (There is no word about ChatgPT GOV yet).
Users with premium plans can choose O3-Mini using the ChatgPT drop-down menu. Free users can click or touch the novel “Reason” button on the chat bar or make ChatGPT “generates” the answer again.
From Friday, O3-Mini will also be available via the API OPENAI interface for selected programmers, but at first it will not have an image analysis. Devows can choose the level of “reasoning” (low, medium or high) to “think harder” based on their cases of operate and the need for delay.
O3-Mini costs 0.55 USD for a million buffished input tokens and USD 4.40 for million production tokens, in which a million tokens correspond to about 750,000 words. This is 63% cheaper than O1-Mini and competitive with the price of the R1 Deepseek reasoning. Deepseek charges $ 0.14 for a million buffished input tokens and USD 2.19 for million output tokens for access to R1 via its API interface.
In chatgpt O3-Mini is set to medium reasoning, which according to OPENAI provides a “balanced compromise between speed and accuracy.” Paid users will be able to choose “O3-Mini-High” in a model that will provide what OpenAI calls a “higher intelligence” in exchange for slower answers.
Regardless of which version of the O3-Mini ChatgPT users, the model will work with the search to find current answers with links to the appropriate online sources. Opeli warns that functionality is a “prototype” because it works to integrate searching in its reasoning models.
“While O1 remains our wider model of general reasoning, O3-Mini provides a specialist alternative to technical fields that require precision and speed,” wrote OpenAI in a blog post on Friday. “The O3-Mini edition means another step in the OpenAI mission to cross the boundaries of profitable intelligence.”
Lots of reservations
O3-Mini is not the most powerful Openai model so far, nor does it jump the R1 Deepseek reasoning model at any reference point.
O3-Mini defeats R1 on Aime 2024, a test that measures how well models understand and react to elaborate instructions-but only with great effort of reasoning. He also overcomes the R1 in a proven verified program program (according to low reasoning O3-Mini delays R1 to Diamond GPQA, which tests models with questions about physics, biology and chemistry at doctoral level.
To be candid, O3-Mini answers many questions at a competitively low cost and delay. In the OpenAI post, he compares its performance with the O1 family:
“With low effort of reasoning O3-Mini reaches comparable performance with O1-Mini, and with an average effort O3-Mini reaches comparable performance with O1,” writes Opeli. “O3-mini with medium reasoning corresponds to the performance of mathematics, coding and sciences, while providing faster answers. Meanwhile, with high reasoning, O3-Mini exceeds both O1-Mini and O1. “
It is worth noting that the advantage of O3-Mini performance over O1 is diminutive in some areas. At Aime 2024 O3-Mini, he overcomes O1 by only 0.3 percentage points when he is set to high reasoning. And on GPQA Diamond O3-Mini does not exceed the O1 result even in the case of high reasoning.
Opeli claims that the O3-Mini is as “safer” or safer than the O1 family, but thanks to the efforts with the red team and its methodology of “alignment”, which makes the models “think” about Openai’s security policy while responding to reacting to the query . According to O3-Mini, it “significantly exceeds” one of the flagship models of OPENAI, GPT-4O, on “difficult safety assessments and jailbreak”.