OpenAI has just released two brand new models, GPT-3 Mini and GPT-4 Mini, which are incredible cutting-edge models. The best part is that they have access to full tool usage, which seems to be a new scaling law.

Not only did we get GPT-3 Mini and GPT-4 Mini, but OpenAI also dropped a brand new project that looks incredible - the Codeex CLI, which is an agentive coding tool powered by OpenAI's models.

According to OpenAI, these are the smartest models they've released to date, representing a step change in ChatGPT's capabilities. The key ingredient is their genetic tool use capabilities, which allow them to use these tools effectively. This is the first time OpenAI has released models with true tool usage out of the gate.

These models are also capable of novel ideas, which is a prerequisite for hitting the intelligence explosion and allowing the models to research and improve themselves iteratively. They are also multimodal, able to input and output different modalities like text, images, and audio.

The benchmarks for these models are impressive, with GPT-3 Mini and GPT-4 Mini performing exceptionally well on tasks like the AMT 2024 and 2025 math competitions, CodeForces coding challenges, and the Humanity's Last Exam benchmark.

The introduction of these models, along with the Codeex CLI tool, represents a significant step forward in OpenAI's efforts to push the boundaries of AI capabilities. However, it's important to be mindful of the platform risk associated with relying too heavily on a single provider like OpenAI, as they may eventually build competing products that could impact your own projects.

The new OpenAI models, GPT-3.5 and GPT-4 Mini, represent a significant step forward in language model capabilities. According to OpenAI, these are the smartest models they have released to date, with a step change in ChatGPT's abilities.

The key advancement highlighted is the models' genetic tool use capabilities. Unlike previous GPT models, these new versions can effectively utilize various tools to accomplish tasks. This includes iterative tool use, where the model tries different approaches to solve a problem, much like a human would.

Another notable feature is the models' ability to generate truly novel ideas, which is a prerequisite for the intelligence explosion. These models are not just regurgitating information but can come up with original solutions and concepts.

Furthermore, the models are multimodal, meaning they can process and generate content in various formats, including text, images, and audio. This allows for more versatile and comprehensive problem-solving capabilities.

Overall, the combination of tool usage, novel idea generation, and multimodality represents a significant advancement in language model capabilities, positioning these new OpenAI models as powerful and flexible AI assistants.

The new OpenAI models, 03 and 04 Mini, have demonstrated impressive performance across a wide range of benchmarks. According to the blog post, these models represent a significant step forward in ChatGPT's capabilities, with the ability to use tools effectively being a key differentiator.

The benchmarks show that 03 and 04 Mini excel in various tasks, including the AMY 2024 and 2025 math competitions, where they outperform the previous 01 model. The models also perform exceptionally well on the CodeForces coding competition, placing them in the top 200 rankings globally.

In the GPQA Diamond benchmark, which tests PhD-level science questions, the new models show a substantial improvement over their predecessors. Similarly, on the Humanity's Last Exam benchmark, 03 with Python and browsing tools achieves a 25% score, a significant jump from the previous 01 and 03 Mini models.

The multimodal benchmarks, such as MMU college-level visual problem-solving and Math Vista, also demonstrate the models' strong performance, with 03 and 04 Mini outscoring the 01 model.

Notably, the Sui Lancer benchmark, which tasks the models with completing real-world software engineering tasks and earning money, shows impressive results. The 03 model achieves a high earnings of $65,000, highlighting its practical capabilities.

Overall, the benchmarks showcase the remarkable progress made by OpenAI in developing these new models, which exhibit enhanced reasoning abilities, tool usage, and multimodal capabilities. The impressive scores across a diverse range of tasks underscore the potential of these models to tackle complex challenges and push the boundaries of what is possible with large language models.

According to the speaker's theory, the new OpenAI models (03, 04 Mini) are the result of a scaling law approach, where they are taking a strong base model (GPT-5) and continuously training it with reinforcement learning. The key insights are:

• OpenAI has likely discovered that they can significantly improve GPT-5 through continued training and refinement.
• They are taking different checkpoints of the evolving GPT-5 model and using those as the starting point for the new 03 and 04 Mini models.
• By applying reinforcement learning with verifiable rewards, they are able to elicit the desired "thinking" behavior and capabilities in these new models.
• This scaling law approach allows for iterative improvements, with each new checkpoint surpassing the previous generation in performance across a wide range of benchmarks.
• The speaker believes this explains the steady, almost "unfettered" progress we are seeing in the capabilities of these new OpenAI models.

In essence, the speaker proposes that OpenAI has found a powerful scaling strategy, continuously building upon and refining their base language model to drive rapid advancements in AI capabilities.

OpenAI's latest models, O3 and O4 Mini, have not only demonstrated impressive performance on various benchmarks, but they have also placed a strong emphasis on cost-efficiency. The company has recognized that as enterprises and developers choose which AI models to build their tools on, cost will be a significant factor in their decision-making process.

The data presented in the transcript shows that the O4 Mini model outperforms its predecessors, such as GPT-4.1, while maintaining a similar or even lower inference cost. This is a significant achievement, as it allows users to access cutting-edge AI capabilities without incurring prohibitive expenses.

Furthermore, the comparison between the O1 and O3 models highlights OpenAI's focus on delivering more cost-effective solutions. The O3 model not only outperforms the O1 model across various benchmarks but does so at a lower inference cost, providing users with a more efficient and affordable option.

This emphasis on cost-efficiency is a strategic move by OpenAI, as it positions their models as attractive choices for a wide range of applications and use cases. By balancing performance and affordability, OpenAI is making their AI technologies more accessible and appealing to a broader audience, from individual developers to large-scale enterprises.

OpenAI has recently launched a new open-source project called Codeex CLI, which is an agentive coding assistant powered by their latest language models. This tool allows developers to leverage the advanced capabilities of OpenAI's models, including multimodal reasoning, tool usage, and the ability to generate novel ideas, directly within their local coding environment.

Codeex CLI can read files from your computer, write files to your computer, and provide a range of coding-related functionalities. This integration of the model's capabilities with your local codebase enables a seamless and powerful coding experience, where the AI assistant can assist with tasks such as code generation, refactoring, and even high-level problem-solving.

One of the key advantages of Codeex CLI is its open-source nature, which allows developers to inspect the underlying code, contribute to its development, and potentially customize it to their specific needs. This approach helps to mitigate the platform risk associated with relying solely on a single model provider, as it gives users more control and flexibility over the tools they use.

Additionally, OpenAI is launching a $1 million initiative to support projects using Codeex CLI and their models. Developers are encouraged to submit proposals for grants in increments of $25,000 in the form of API credits, which can help to offset the costs of using these advanced AI capabilities.

Overall, Codeex CLI represents an exciting development in the field of AI-powered coding assistants, offering developers a powerful and open-source tool to leverage the latest advancements in language models and agentive reasoning within their local development workflows.

While the new OpenAI models, 03 and 04 Mini, represent significant advancements in AI capabilities, particularly in terms of tool usage and novel idea generation, there are important considerations to keep in mind.

The platform risk associated with building on top of OpenAI's models is a valid concern. As the company continues to expand its offerings, there is a risk that they may encroach on the markets and projects of their own customers. This highlights the importance of diversifying one's AI model dependencies and exploring open-source alternatives.

However, the cost-efficiency and performance improvements offered by these new models are undeniably compelling. Developers and enterprises will need to carefully weigh the benefits against the potential risks when deciding which models to incorporate into their projects.

Additionally, OpenAI's $1 million initiative to support projects using Codeex CLI and their models presents an opportunity for innovative developers to access valuable resources and funding. This could help mitigate the platform risk by fostering a more diverse ecosystem of AI-powered applications.

In the end, navigating the evolving AI landscape requires a balanced approach. Staying informed, diversifying dependencies, and exploring open-source options can help organizations and developers capitalize on the remarkable advancements in AI while managing the inherent risks.