Photo credit: arstechnica.com

OpenAI Unveils New Models o3 and o4-mini with Enhanced Capabilities

On Wednesday, OpenAI introduced its latest advancements in artificial intelligence with the launch of two new models: o3 and o4-mini. These models stand out for their ability to integrate simulated reasoning skills with functionalities such as web browsing and coding support. This is the first instance where OpenAI’s models focusing on reasoning can utilize all ChatGPT tools at once, which includes visual analysis and image generation.

OpenAI first revealed model o3 back in December. Prior to this launch, only less advanced variants known as “o3-mini” and “o3-mini-high” were available. The introduction of o3 and o4-mini effectively phases out the earlier versions—o1 and o3-mini.

Starting today, ChatGPT Plus, Pro, and Team users will have access to these new models, while Enterprise and Edu users will receive access in the following week. Free users can experiment with o4-mini through the “Think” option available before submitting their queries. OpenAI’s CEO, Sam Altman, also tweeted that “we expect to release o3-pro to the pro tier in a few weeks.”

For developers, o3 and o4-mini are now accessible via the Chat Completions API and Responses API, although some organizations might need to undergo a verification process to gain access.

According to OpenAI, “These are the smartest models we’ve released to date, representing a significant leap in the capabilities of ChatGPT for a wide range of users, from curious individuals to advanced scholars.” It has been noted that these models are more cost-efficient compared to their predecessors. Each model serves a distinct purpose: o3 is tailored for complex analytical tasks, while o4-mini, being a smaller variant of the forthcoming high-performance model “o4,” is optimized for speed and cost-effectiveness.

OpenAI emphasizes that o3 and o4-mini are multimodal, describing their ability to “think with images.”

One of the defining features of these new models is their simulated reasoning capability, which employs a step-by-step “thinking” approach to address problems. Moreover, these models can autonomously decide when and how to use support tools for tackling intricate challenges. For instance, if prompted about projected energy consumption in California, the models can independently search for relevant utility data, write Python code to generate forecasts, create visual representations, and elucidate the critical variables influencing their predictions—all within the context of a single query.

Source
arstechnica.com