OpenAI launches o3-mini, its latest ‘reasoning’ model

OpenAI launched a new AI “logic” model on Friday, and 3-minute, new in the company O family of logic modelsThe

Open The first December of the model in December In addition to a more capable system called O3, however, this introduction comes at an important moment for the organization, whose ambitions – and challenges – apparently the day is increasing.

Opina is fighting the idea that it is in the AI ​​race DIPSECWhich Open has complained that it has stolen the IP. It has been tried On the shore of its relationship with Washington As it follows a one at the same time Ambitious data center projectAnd As it gives it the foundation of the foundation For the biggest funds in history.

Which brings us to 3-minute. Open is pitching its new model as “strong” and “affordable”.

“The sign of today’s launch […] An important step towards the expansion of accessibility in advanced AI in our mission services, “a spokesperson for OpenAI told TechCrunch.

More efficient argument

In contrast to most large language models, reasonable models such as O3-Mini do the results perfectly before giving them results. It helps them Avoid some problems That usually up the model trip. These rational models take a little longer to come to the solutions, but the trade is closed, but they are more reliable in domains like physics-even if not perfect.

O3-Mini Stem Problems, especially for programming, mathematics and fine for science. OpenAI has claimed that the model is basically equal to 1 family, and 1 and 1-mint of skill, but run fast and spends less.

The agency claims that external examiners and-3-min-minimum answers and more than 1-minute are more than half the time. O3-Mini obviously created 39% less “big mistakes” “stiff real-world questions” A/B exam Vs. O1-Mini, and about 24% quickly makes a “clear” response while answering quickly.

Will be available for all users through O3-Mini Chatzipt Starting from Friday, however, the OPGPT ChatGPT Plus and teammates will get a higher rate limit of 150 questions per day. ChatzPT Pro customers will receive unlimited access and O3-Mini will come to ChatzPT Enterprise and ChatzPT ADU customers in a week. (No words Chatzipt gov Still).

Premium planned users can select O3-Mini using the ChatzPT Drop-Down menu. Users can click or tap the new “cause” button in the chat bar, or “re -generate” an answer to the chatter.

Starting from Friday, and the 3-minute Openai will be available to select developers through the API, but there will be no support for analyzing images initially. Daves can choose the level of “logic effort” (lower, medium or high) to get O3-Mini to “more difficult thinking” on the basis of their use case and delayed need.

The price of O3-Mini is 40.40 per million Cashed input token per million input tokens, where one million token is equal to 750,000 words. It is 63% cheaper than O1-Mini and competitive with DEPSEC R1 argument model pricing. DEPSEC charges $ 2.19 per million input token per million input tokens and millions of output token for 1 access through API.

In the ChatzPT, O3-Mini is set in a moderately rational effort, which Openi says “a balanced trade between speed and accuracy” provides. Paid users will have the option of selecting “O3-Mini-High” on the model picker, which will supply “higher intelligence” in exchange for slow reactions.

O3-Mini ChatgPT does not choose which version of users, the model will work with the links to find up-to-date answers with links to relevant web sources. OpenAI warns that functionality is a “prototype” because it works to integrate the search on its rational models.

Opina wrote in a blog post on Friday, “Although Our 1 remains our broad general-knowledge argument model, and provides a special alternative for precision for 3-minute technical domains,” OPENA wrote in a blog post on Friday. “And the release of the 3-minute release identifies another step to push the OpenAE mission to the boundaries of expensive intelligence.”

Plenty of caution

O3-Mini is not the most powerful model of the Openi till today, or it does not lip the DIPSC’s RA logic model in every benchmark.

O3-Mini beats R1 in AIM 2024 It also beats the SWU-Bench (by .1 points) in programming-centric examination (by .1 points), but again with high rational efforts. In a short argument effort, GPQA backs the Diamond and 3-Mint R1 lags, which examine models with PhD-level physics, biology and chemistry questions.

Frankly, and 3-minute competitively answered many questions at low cost and delay. In the post, compares OpenAI’s performance to the O1 family:

Opina writes, “With less reasonable efforts, and achieved comparative performance with 3-minute and 1-minute, when he achieved medium efforts and comparative performances with 3-minute and 1,” Written by Open. “Mathematics, coding and matching with a 3-minute fast response with a medium rational effort and 3-minute performance.

It is noteworthy that the performance benefits of 3-minute and 3-minute performance in some cases are slim. At 2024 AIM, and when set in a 3-minute high reasoning attempt and defeated 1 by only 0.3 percent point. And in GPQA Diamond, and 3-minute high reasoning efforts do not even exceed the score of 1.

Opina emphasizes that he is “safe” or safe than 3-minute and 1 families, however, thanks to the Red-Timing attempt and its “deliberate alignment” method, which makes the models “thought” about Openi’s protection policy when they react Question. According to the agency, one of the flagship models of the 3-Minnet OpenAEE “has exceeded significantly”, GPT -4O“Challenging Protection and Jailbreak assessment.”

There is an AI-centric newsletter on TechCrunch! Sign up here Every Wednesday gets it to your inbox.