AiNews.com
Posts
OpenAI o3-mini: Faster, Smarter AI for STEM, Now Free for ChatGPT Users

OpenAI o3-mini: Faster, Smarter AI for STEM, Now Free for ChatGPT Users

Alicia Shapiro
February 03, 2025 • Estimated Reading Time: 7 minutes

Image Source: ChatGPT-4o

OpenAI o3-mini: Faster, Smarter AI for STEM, Now Free for ChatGPT Users

OpenAI has officially released o3-mini, the latest addition to its cost-efficient reasoning model series, available now in both ChatGPT and the API. First previewed in December 2024, this model pushes the boundaries of small-scale AI, excelling in STEM fields—science, math, and coding—while maintaining low costs and reduced latency similar to its predecessor, OpenAI o1-mini.

Key Features and Improvements

Developer-Focused Enhancements

OpenAI o3-mini is the first small reasoning model to support function calling, structured outputs, and developer messages, making it ready for production use from day one. It also supports streaming and offers three reasoning effort options—low, medium, and high—allowing users to prioritize either speed or deeper analysis for complex tasks.

Wider Access and Increased Limits

Available immediately for ChatGPT Plus, Team, and Pro users, with Enterprise access coming later in February.
o3-mini replaces o1-mini in the model picker, offering higher rate limits and lower latency.
Plus and Team users now have triple the daily message limit—150 messages per day instead of 50.
o3-mini now includes an early search prototype, allowing it to find up-to-date answers with links to relevant web sources.
Future integrations planned to expand search capabilities across OpenAI’s reasoning models.
Free users can also try o3-mini by selecting ‘Reason’ in the ChatGPT message composer—the first time OpenAI has made a reasoning model available at no cost.

A table comparing general knowledge performance across OpenAI models, evaluating accuracy in general reasoning (MMLU), math (Math and MGSM), and factuality (SimpleQA). The table shows o3-mini outperforming o1-mini in multiple categories, with higher accuracy in math and factuality, especially at medium and high reasoning effort levels.

General Knowledge Performance Across AI Models. Image Source OpenAI

Enhanced Performance in STEM and Logic

OpenAI claims that o3-mini is significantly better than o1-mini in math, coding, and scientific reasoning. In testing:

Users preferred o3-mini’s answers 56% of the time over o1-mini.
It reduced major errors by 39% on difficult real-world questions.
With medium reasoning effort, o3-mini matches o1’s performance in competition-level math and PhD-level science evaluations, including AIME and GPQA.
In coding benchmarks (Codeforces, SWE-bench, LiveBench), o3-mini outperforms o1-mini and, at higher reasoning effort, surpasses o1.

A bar chart comparing accuracy on AIME 2024 competition math questions across OpenAI models. The chart shows o3-mini achieving higher accuracy than o1-mini at all reasoning effort levels. The highest accuracy, 87.3%, is reached with o3-mini (high), surpassing both o1-mini and o1. The gray-shaded regions indicate majority vote performance with 64 samples.

Competition Math (AIME 2024) Performance. Image Source: OpenAI

A bar chart comparing accuracy on software engineering tasks (SWE-bench Verified) across OpenAI models. The graph shows o3-mini surpassing o1-mini at all reasoning levels, with the highest accuracy (49.3%) achieved by o3-mini (high). The chart highlights o3-mini as OpenAI’s best-performing released model for software engineering tasks.

Software Engineering (SWE-bench Verified) Accuracy. Image Source: OpenAI

Speed and Efficiency Gains

Compared to o1-mini, o3-mini:

Delivers responses 24% faster—average 7.7 seconds compared to 10.16 seconds.
Has 2500ms lower latency on the first token generation.
Offers an early prototype of a search function, allowing it to fetch up-to-date answers with relevant web links.

Safety and Reliability

To enhance AI safety, OpenAI used deliberative alignment, a method where the model reasons about human-written safety guidelines before responding. Evaluations suggest o3-mini significantly performs better than GPT-4o in safety and jailbreak tests, and OpenAI conducted rigorous risk assessments including external red-teaming and safety evaluations before release.

A screenshot of the OpenAI o3-mini System Card displaying key risk areas and a preparedness scorecard. The risk section highlights concerns related to disallowed content, jailbreaks, and hallucinations, all marked with checkmarks. The preparedness scorecard evaluates o3-mini’s readiness in CBRN (medium), cybersecurity (low), persuasion (medium), and model autonomy (medium), using a visual bar indicator to represent different risk levels.

OpenAI o3-mini System Card: Risk and Preparedness Evaluation. Image Source: OpenAI

What This Means

With o3-mini, OpenAI continues its push for more affordable and efficient AI models, making advanced reasoning capabilities more accessible. By tripling message limits, expanding free access, and improving speed and accuracy, OpenAI is strengthening its position in AI-driven STEM applications.

As AI competition heats up—particularly with challenges from models like DeepSeek—OpenAI's latest release shows its commitment to affordable intelligence without compromising performance. Looking ahead, further optimizations and integrations, especially in search functionality, could make models like o3-mini even more valuable for developers and general users alike.

To see more benchmarks, please visit OpenAI’s announcement, and visit the system card to learn more details.

Editor’s Note: This article was created by Alicia Shapiro, CMO of AiNews.com, with writing, image, and idea-generation support from ChatGPT, an AI assistant. However, the final perspective and editorial choices are solely Alicia Shapiro’s. Special thanks to ChatGPT for assistance with research and editorial support in crafting this article.