On April 16, 2025, the OpenAI team hosted a Question and Answer (AMA) event on Platform X, detailing the latest advances and technical details of their upcoming release.
In terms of product releases and roadmap, OpenAI announced that the reinforcement fine-tuning feature will soon be fully open to support the public in fine-tuning inference models or reinforcement learning.furthermore, OpenAI plans to release a robust open source model in the coming months and welcomes feedback from developers to improve its utility.
Regarding the image generation functionality, new features in the API will be available soon. o3-mini will not support visual processing in the API, but o4-mini already supports visual inputs and is currently available as an inference mini-model.More advanced o3-pro modelsIt is in development and is expected to be released soon.
In terms of model performance and usage, the entire GPT-4.1 family was trained specifically for coding and agent tasks. o3 is suitable for complex coding problems or broad topic searches, while o4-mini is more suited for fast interactions or smaller coding tasks. o4-mini has a higher rate limitation, and performs as well as o3 in evaluations involving vision, math, science, and some coding tasks equivalent.For tasks requiring in-depth knowledge and nuance, o3 still has a clear advantage.
GPT-4.1 can improve performance in specific scenarios and reduce the number of examples in the cue through customized fine-tuning.GPT-4.1 training is used for exploring possible solutions, self-reflection, and self-correction.o3 and o4-miniBoth provide a context window of 200k tokens with a maximum output token count of 100k and have been evaluated for cybersecurity both internally and by third parties.
In terms of API features and tool support, the Response API can automatically switch between developer and system messages. A parameter called "reasoning effort" can be used to control costs, especially as o-series models become more intelligent.at the present time, the ChatCompletions or Responses APIs do not support hosted tools, but once support goes live, developers will be able to control what tools are enabled.
The SDK automatically converts Pydantic schema to JSON-compatible schema, handling properties such as "additionalProperties".Tools such as web search, file search and code interpretersIt is actively used in the inference phase of o3 and o4-mini, and although these tools are currently only supported in ChatGPT, API support is coming soon.
The Agent SDK and Codex CLI are also under active development.Threading SupportConversation history and retention will be improved.The Codex CLI includes multiple documented approval modes that allow users to select modes based on action or session.Codex CLIDesigned to run background tasks, not a replacement for IDE tools.
---
OpenAI has announced the release of two powerful new inference models in its API: o3 and o4-mini. these models excel in multiple domains and provide developers with more options to choose from.
The o3 model achieves industry-leading performance in programming, math, science, and visual processing for application scenarios that require high precision and complex reasoning. For example, in software development, o3 can help programmers automatically generate code snippets to improve development efficiency; in scientific research, it can assist researchers in data analysis and model construction.
o4-mini On the other hand, it is a faster and more cost-effective inference model for scenarios where response time and cost are more important. Although it performs slightly less well than o3 on some tasks, the efficiency and affordability of o4-mini makes it ideal for many real-world applications.
The introduction of these new models further solidifies OpenAI's leadership in AI and provides developers across industries with additional tools to enable more efficient and smarter solutions.
---
OpenAI recently released three new inference models o3, o4-mini and o4-mini-high, codenamed Phoenix. these models are designed to improve the performance and functionality of ChatGPT.
o3 replaces the previous version o1, while o4-mini and o4-mini-high replace o3-mini and o3-mini-high, respectively. These new models support a wide range of ChatGPT tools, such as Python code execution, search, canvas manipulation, file uploading, image analysis, and image generation. In addition, they provide detailed "structured thinking", showing explicit reasoning steps, search results, Python code, and collapsible summaries. It is worth noting that these models perform well on image processing and advanced visual reasoning tasks, achieving top benchmark scores.
These new models offer significant improvements in accuracy, efficiency, and cost-effectiveness, and set new industry records in coding, math and vision tasks. To ensure safety, these models utilize the latest rejection training techniques, advanced risk detection mechanisms, and are independently reviewed to ensure that they consistently fall below all high-risk safety thresholds.
ChatGPT Plus, Pro and Team users now have access to these new models, Enterprise and Edu users will get access within a week. Free users can try out o4-mini via Think mode. o3 and o4-mini are already supported by the API and Responses API for advanced reasoning and to provide a summary of the reasoning (tool support will be available in the coming weeks).
- In the API, o4-mini and o3 provide a context window of 200,000 tokens and support up to 100,000 output tokens.
- The knowledge deadline is June 1, 2024.
In addition, Codex CLI is now being released as an open source endpoint coding agent that brings ChatGPT-level code inference, real-time code execution, multimodal support, and security sandbox automation to the endpoint.OpenAI has also set up a $1 million grant program for this purpose to support open source projects that use the Codex CLI and OpenAI models.
---
Veo 2 is now live on Google AI Studio and the Gemini API, a powerful video generation tool that will allow users to easily transform text and image cues into stunning video content.
By integrating Veo 2, developers can dramatically improve the efficiency and quality of their video production by seamlessly incorporating it into their existing workflows.Utilizing advanced deep learning technology, Veo 2 is capable of generating high-quality videos based on simple text descriptions or graphic inputs for a wide range of applications, including advertising, education, entertainment, and more.
Application Scenarios:
- **Advertising Ideas**: Quickly generate eye-catching advertising videos to increase brand exposure.
- **Online education**: visualizing course content to enhance the learning experience.
- **Social Media**: create compelling video content that attracts more attention.
Learn more about how to use Veo 2 and its features at: https://t.co/jDff8aOTkS
---
The latest version of the AI SDK now integrates OpenAI's newest inference models, o3 and o4-mini, which excel at complex inference tasks and provide more accurate and efficient solutions.
The o3 model is designed for application scenarios that require high accuracy and powerful inference capabilities, and is suitable for natural language processing, data analysis and intelligent decision making. The o4-mini, on the other hand, is a lightweight model suitable for resource-constrained environments such as mobile devices and edge computing.
Key features include:
- Significantly improved inference performance
- Higher accuracy and reliability
- Optimized resource efficiency
Developers can visit the official website for more detailed information and technical documentation to quickly get started and take full advantage of these advanced models.
---
Codex CLI is an open source native programming agent that converts natural language into runnable code. Just tell Codex CLI what you need to build, fix, or interpret, and it will quickly bring your ideas to life.
This tool is especially suited for developers and beginners to significantly improve coding efficiency and reduce the learning curve.Through a simple command line interfaceThe Codex CLI allows users to easily interact with the Codex CLI and get instant feedback.
The Codex CLI is used in a wide range of scenarios, including but not limited to:
- Rapid Prototyping
- Automated Script Generation
- Code debugging and fixing
In addition, the Codex CLI supports a wide range of programming languages, such as Python, JavaScript, and Java, making it an indispensable tool for multilingual developers.
Whether you're an experienced programmer or just starting out, the Codex CLI can help you accomplish your programming tasks more efficiently.
---
OpenAI today (April 16, 2025) officially released its newest flagship model for "reasoning"o3and its Lite versiono4-minio3 is positioned as the most powerful multimodal reasoning model available, autonomously invoking browser, Python code, file parsing, and image generation in less than a minute, and for the first time integrating "think in pictures" into the reasoning chain, while o4-mini delivers superior mathematical and programmatic performance at low latency and low cost. Both models are available in ChatGPT Plus, Pro, Team, and API, and have completed a security review through a new version of the Preparedness framework.
The o-series focuses on "deep thinking" rather than pure language generation. o3 was announced in preview on December 20, 2024, with a name that avoided conflict with telecom operator O2; then on January 31, 2025, the adult version o3-mini was launched; and the full version of o3 is fully open today. In terms of core technology upgrades, theSimulated reasoningThe mechanism allows the model to conduct "private chain reflection" before answering, which significantly reduces the illusion; "deep reinforcement learning extension" verifies the quadratic scaling of "increased arithmetic power → increased inference ability"; "deep reinforcement learning extension" verifies the quadratic scaling of "increased arithmetic power → increased inference ability". Official tests show that o3 can be used in the real world. Official tests show that o3 has 20% fewer major errors than o1 on real-world problems.
o3 and o4-mini can embed images in chained thinking, parse fuzzy boards, rotate/zoom schematics, and carry visual insights into the final answer, opening up a new paradigm of hybrid text-visual reasoning. The model evaluates when to retrieve a web page, write Python code, or generate an image, and weaves the results into a one-time response, marking ChatGPT's move towards an "agent-based" assistant. In addition, o3 sets Codeforces, SWE-bench, and multimodal MMMU records; and o4-mini achieves class-leading performance at lower cost on polytechnic benchmarks such as AIME 2024/2025 and GPQA Diamond.
In terms of pricing, o3 is priced at $10.00 per million Token for inputs and $40.00 per million Token for outputs for high-precision multimodal reasoning, while o4-mini is priced at $1.10 and $4.40, respectively, for fast and low-cost STEM and code tasks. Pricing has dropped significantly from its o1 counterpart, the ChatGPT payment tier can be switched directly from the model selector, and the enterprise version of o3-pro will be live in a few weeks.
In terms of developer ecology, the o3 series supports Chat Completions, Assistants, Batch APIs, and natively provides function calls, structured output and developer messages. The Codex CLI, an open source endpoint agent, can be invoked offline to manipulate native code workflows with o3/o4-mini. In addition, the o3-powered web research agent automatically collects and aggregates data, and is considered a model for the AI agent era.
While OpenAI's release of o3 ahead of competitors such as Google Gemini 2.0 and DeepSeek has shortened the security testing cycle and sparked discussion about its adequacy, industry observers point out that o3's multimodal and agent characteristics may reshape data analysis, scientific research and software development workflows, and that the future will also place higher demands on model transparency and alignment.
---
Update: Starting today, two new models, o3 and o4-mini, will be officially available on the ChatGPT and API platforms.
These two models have been specially trained to have more powerful tool usage capabilities. This update marks a further breakthrough in the practical application of AI technology, which is expected to significantly improve user experience and work efficiency.
Key Features:
- **Enhanced tool usage**: o3 and o4-mini are able to better understand and manipulate tools to provide more precise and practical answers.
- **Broader application scenarios**: these models will be applicable to a wider range of domains, such as customer service, technical support and content creation.
- **Higher Flexibility**: Through API interfaces, developers can easily integrate these new models for customized solutions.
The update not only brings more choices for users, but also provides more powerful AI support for enterprises, promoting the intelligent process of the whole industry.
---
We are pleased to announce the launch of the new /extract v2, a revolutionary data extraction tool that provides users with more power and flexibility.
The new version of /extract v2 supports extracting data from multiple pages and handles paging, and can even extract data from the entire web without a specific URL. This technological breakthrough is driven by FIRE-1, which makes data extraction more efficient and convenient.
The core strength of **/extract v2** is its advanced endpoint design, which allows users to access the data they need with simple prompts. This technology was significantly improved on the third day of Launch Week, providing users with even more powerful data extraction capabilities.
Here are the main features of /extract v2:
- Multi-page data extraction
- Automatic handling of page breaks
- URL-less full web data crawling
These features make /extract v2 an ideal tool for data scientists, researchers and developers in a wide range of applications such as market analysis, competitive intelligence, academic research and many other fields.
---
The latest news is that the o3 and o4-mini tools are now officially available on the Playground platform. These two tools are an important part of the AI development space, aiming to provide developers with a more efficient and convenient development experience.
o3 is a powerful machine learning model training tool that supports a variety of deep learning frameworks and can significantly improve the speed and accuracy of model training. And o4-mini is a lightweight inference engine designed for edge computing and mobile devices with low power consumption and high performance.
By using o3 and o4-mini on Playground, developers can easily train and deploy models, accelerating projects from concept to real-world application. In addition, Playground provides a wealth of sample code and documentation to help developers get started quickly.
Whether you're a beginner or an experienced developer, you'll find the right tools and technical resources for you on the Playground platform.
