Influenced by DeepSeek? OpenAI goes live with O3 Mini, inference models free for the first time

Friday, January 31st, local time.OpenAIThe official launch of a new inference modelo3-mini, and for the first time, the inference model has been made available to free users. This is the newest and most cost-effective model in OpenAI's inference family, which OpenAI says is capable of human-like reasoning thatIt's now live in ChatGPT and the API. With theDeepSeekReleasing an open source model that shook the world, this new product from OpenAI is getting a lot of attention.

The o3-mini provides users with STEM capabilities at a lower cost and faster response time, particularly in the areas of science, math, and programming, while continuing the low-cost and low-latency features of previous versions such as the o1-mini. It is worth noting that developers can choose between three different "reasoning effort" options, low, medium and high, depending on their needs.

OpenAI says that while OpenAI o1 remains the broad model for general knowledge reasoning, OpenAI o3-mini provides a specialized alternative for technical areas where accuracy and fast response are required. o3-mini balances speed with accuracy by using moderate reasoning effort.

Support for more developer features

o3-mini is the first small inference model that supports features commonly used by developers, included:

Function call: You can directly call the preset functions to realize more efficient task processing.

Structured Output: make the information in the model output more regular and easy to parse and apply.

Developer messages: provide developers with more debugging and information feedback means.

Like the previous o1-mini, the latest o3-mini also supports streaming output.

OpenAI describes thato3-mini supports developers to adjust the "AI reasoning effort", which is categorized into three levels: low, medium and high.This flexibility allows o3-mini to "think harder" when faced with difficult problems, while prioritizing speed of response when efficiency is required.

The o3-mini does not support visual capabilities, so for users who need to perform image processing or visual reasoning tasks, they still need to use OpenAI's o1 model.

Wide range of access options

For different types of users, o3-mini can be accessed through multiple channels:

• API Users: o3-mini is already available to some API users in the Chat Completions, Assistants, and Batch APIs (for users with usage tiers 3-5).
• ChatGPT Users: ChatGPT Plus, Team and Pro users will be able to use it starting Friday, while enterprise users will get access a week later.
• Free users: Free users can also experience o3-mini by selecting the "Reasoning Mode" or by regenerating their answers. this is the first time that free users have access to a model with reasoning capabilities.

OpenAI says that o3-mini will replace OpenAI o1-mini in the model selector.All paid subscribers can choose o3-mini-high in the model selector - a version with more intelligence, but slightly slower generation response.

Pro users have unlimited access to o3-mini and o3-mini-high.OpenAI increased the daily message limit for Plus and Team users from 50 for o1-mini to 150 for o3-mini.

In addition.o3-mini now supports search functionthat can find up-to-date answers and provide links to relevant web pages. This is an early prototype, and OpenAI says it is working to integrate the search function into all of its inference models.

Model Performance Highlights

According to OpenAI's disclosure, in the tests of the American Invitational Mathematics Tournament 2024 (AIME 2024), the accuracy of o3-mini at low reasoning effort was 60%, which is about the same as that of o1-mini, but faster; while at medium effort, o3-mini was able to improve its accuracy to 79.6%, which is comparable to that of the o1 model; and at the highest level of effort whenThe accuracy of o3-mini can be further improved to 87.3%.

For the doctoral level scientific questions (GPQA Diamond), the accuracy of the three effort level models was 70.61 TP4T, 76.81 TP4T, and 79.71 TP4T, respectively.

FrontierMath Frontier Math and Codeforces and other programming competitions, o3-mini also shows a clear advantage, even far surpassing its predecessor model in some reviews.

In the SWE-bench Verified software engineering task test, the o3-mini high inference version achieves an accuracy of more than 49%, outperforming the older version.

In terms of general knowledge, the o3-mini also outperforms the o1-mini in a variety of knowledge reviews and is able to provide users with more accurate answers.

With a level of intelligence comparable to that of the o1, the o3-mini offers faster performance and greater efficiency.In addition to the STEM assessments mentioned above, using moderate reasoning effort, the o3-mini also demonstrated better performance on the math and factual assessments. In the A/B test, the o3-mini responded 241 TP4T faster than the o1-mini, with an average response time of 7.7 seconds compared to 10.16 seconds for the o1-mini. In terms of latency, the first token of o3-mini is on average 2500 ms faster than o1-mini.

Security and Risk Prevention and Control

OpenAI says that o3-mini uses a "deliberate alignment" approach, which allows the model to think about human-created security rules before answering user questions. Similar to the o1 model, o3-mini outperforms GPT-4o in addressing complex security challenges and preventing jailbreaks.

Prior to the release, OpenAI rigorously assessed the risks of o3-mini using comprehensive security preparation, external red team testing, and multiple security assessment methods. The relevant detailed assessment results and risk prevention and control measures are recorded in the system card of o3-mini.

future outlook

The release of the o3-mini marks another step in OpenAI's push to push the boundaries of low-cost intelligence, OpenAI said.Making high-quality AI more pervasive by optimizing reasoning power for STEM fields while keeping costs low.OpenAI notes that the model continues a tradition of continually lowering the cost of intelligence - since the launch of GPT-4, the pricing per token has been reduced by 951 TP4T -- while still maintaining top-notch reasoning capabilities.

OpenAI says it will continue to be at the forefront of building large-scale models that balance intelligence, efficiency and safety as AI becomes more widely used.

o3-mini on the eve of its release

The background to the release of the o3-mini is quite striking.

The Trump administration's massive Stargate AI funding program comes just one day after OpenAI announced the Operator AI agent.

Subsequently, the rise of DeepSeek R1 shocked the world and impacted the market, and competition in the AI field intensified, OpenAI accelerated the o3-mini release process to maintain its leadership in the AI field. Even before the official release of o3-mini, there was news that OpenAI was preparing to release this Friday a new generation of inference model, ChatGPT o3-mini, which is a streamlined version of the o3 series, optimized for specific tasks, faster and more cost-effective.

OpenAI CEO Sam Altman said on social media platform X on January 17 that the final version of the ChatGPT o3-mini is complete and is in the process of being released. At the time, he expected the new version to be available "in about a couple weeks."

This article is fromWeChat "Hard AI"For more information about AI, please visitMove here.