OpenAI unveiled two significant releases within the last week: o3-mini, their newest cost-efficient reasoning model, and deep research, an agent capable of comprehensive online research and analysis.
o3-mini: Pushing Cost-Effective Reasoning
While smaller than the upcoming full o3 model, o3-mini demonstrates impressive capabilities across STEM fields. The model achieves 87.3% accuracy on AIME 2024 mathematics competition questions at high reasoning effort, showcasing strong mathematical reasoning abilities. For context, this represents significant progress over previous mini models while maintaining lower computational costs. In some use cases, o3-mini-high even outperforms the current full o1 model.
In PhD-level science evaluation (GPQA Diamond), o3-mini reaches 79.7% accuracy at high reasoning effort, approaching the performance of larger models. The model also shows remarkable coding capabilities, achieving a 2073 Elo rating on Codeforces competitive programming tasks and setting new records on software engineering benchmarks like SWE-bench Verified (49.3% accuracy).
One of o3-mini’s key innovations is its flexible reasoning effort settings—low, medium, and high—allowing developers to optimize for either speed or accuracy. The model delivers responses 24% faster than its predecessor o1-mini, with an average response time of 7.7 seconds compared to 10.16 seconds.
Another notable feature is that o3-mini has a web-enabled option, which is the first time ChatGPT has combined a reasoning model with real-time web search.
Deep Research: A New Frontier in AI Investigation
Alongside o3-mini, OpenAI introduced deep research, an agent that transforms how AI handles complex research tasks. In internal evaluations, deep research demonstrated the ability to automate multiple hours of difficult manual investigation across various domains.
The agent shows particular strength on GAIA, a benchmark for real-world questions requiring reasoning and web browsing capabilities. Deep research achieved a new state-of-the-art score of 67.36% (pass@1) and 72.57% (cons@64), surpassing previous records.
What sets deep research apart is its ability to conduct multi-step research tasks, analyzing hundreds of online sources to create comprehensive reports at the level of a research analyst. The agent can process large amounts of text, images, and PDFs, adapting its search strategy based on discovered information.
Accessibility and Future Plans
Both releases are being rolled out gradually. o3-mini is available today to ChatGPT Plus, Team, and Pro users, with Enterprise access coming in February. Notably, this marks the first time a reasoning model has been made available to free users in ChatGPT.
Deep research is currently available to Pro users, with Plus and Team access planned for the near future. OpenAI is working on expanding its capabilities to include specialized data sources and subscription-based resources.
Looking Ahead
These releases represent significant steps in making advanced AI capabilities more accessible and cost-effective. With o3-mini demonstrating strong performance at lower computational costs and deep research automating complex research tasks, we’re seeing AI tools become both more powerful and more practical for everyday use at an ever accelerating pace.
Benchmark data and full information from OpenAI at the following links:
