Developer State Of The Union

The developer experience is being fundamentally reshaped by AI. OpenAI’s journey, from its founding in 2015 with a three-step plan—solve reinforcement learning, solve unsupervised learning, and gradually learn complicated things—has led to today’s powerful developer platform. Early successes like the Dota 2 AI in 2017 demonstrated the power of reinforcement learning at scale, while the “unsupervised sentiment neuron” paper from the same year proved that next-step prediction could lead to an emergent understanding of semantics. This insight was the catalyst for scaling up, culminating in models like GPT-3.

Instead of building a single application, OpenAI chose to release its technology as an API, empowering developers to connect the models to real-world problems. This developer-first approach has been validated by the explosive growth and innovation seen across the ecosystem.

New and Updated Models

A suite of new and updated models is now available, designed to provide more power, flexibility, and efficiency.

• GPT-5 and GPT-5 Pro: The latest flagship model, GPT-5, is specifically designed for agentic tasks, excelling at instruction following and tool use. It possesses advanced coding intelligence, capable of refactoring large codebases and generating UIs from prompts. To get the most out of GPT-5, developers should follow three principles:

  1. Keep prompts simple: Shorter, clearer instructions yield better results.
  2. Use Evals: A new prompt optimizer is available to help refine prompts.
  3. Iterate systematically: Change only one variable at a time when testing. GPT-5 Pro, the model powering ChatGPT, is now available in the API. It thinks for longer and offers unparalleled intelligence for use cases where precision is critical.

• Sora 2 in the API: The highly anticipated text-to-video model, Sora 2, is now in the API, along with Sora 2 Pro. Sora 2 is ideal for rapid experimentation and prompt refinement, while Sora 2 Pro provides higher fidelity and closer attention to detail for production-quality video.

• Efficient Voice and Image Models: To make AI more accessible, new smaller models for voice and image generation are being released at a fraction of the cost. A new mini real-time voice model offers the same quality as the large model at a 70% lower cost. A new image generation model, already used in ChatGPT, provides similar quality to the original at an 80% price reduction.

• GPT-OSS: The open models initiative has been a massive success, with over 23 million downloads. It provides developers the flexibility to build locally, offline, or on-premise, enabling applications like local security operation centers and mesh networks for disaster relief.

Codex: The Future of Coding

Codex has evolved into a comprehensive coding assistant that works across the terminal, IDE, GitHub, and in the cloud. Powered by a specialized version of GPT-5, it excels at real-time collaboration and autonomous complex work.

Key new features for teams include:

• Slack Integration: Assign tasks to Codex directly from Slack conversations, allowing it to pick up context and create tasks.
• Multimodal Capabilities & MCP Support: Codex can now interpret visual inputs, like whiteboard photos, and generate multiple UI versions to choose from (“best of n”). It supports the Meta-Call Protocol (MCP), enabling integration with tools like Figma (to fetch component code directly from designs) and Chrome DevTools (to analyze front-end performance and suggest improvements).
• Automated Code Reviews: When enabled on a GitHub repository, Codex will automatically review every pull request. It goes beyond static analysis to deeply understand the project, run commands, and catch potential bugs before they reach production.
• Codex SDK: The same powerful agent used in the official tools is now available as an SDK. This allows developers to embed Codex into custom workflows, CI/CD pipelines (via a new GitHub Action), or even directly into applications to create self-evolving software.

Building Agents with the Responses API and Agent Kit

The API for building agents has evolved from simple completions to the industry-standard chat completions. Today, the Responses API is the flagship primitive for building modern, multi-modal agents. It is designed for models that call multiple tools and use chain-of-thought reasoning. It offers key advantages:

• Built-in tools like web search and file search.
• Ability to call multiple tools in a single request.
• Preservation of reasoning tokens across turns, making it ~20% faster on average for multi-turn requests.

Building agents is complex, which is why OpenAI introduced Agent Kit, a set of customizable building blocks built on the Responses API. The kit includes:

• Agent Builder: A visual canvas with templates to compose and orchestrate complex agentic workflows.
• Chatkit: A component to easily deploy a fully-featured chat interface that matches your brand’s look and feel.
• Widget Builder: A tool to create custom, interactive UI components (widgets) that agents can render in the chat.
• Evals and Grading: A massively updated Evals product helps automatically improve prompts and monitor agent performance. You can create custom “graders” to assess agent behavior against specific guidelines.
• Reinforcement Fine-Tuning (RFT): RFT has been extended to tool calls, enabling models to learn to call the right tools at the right time more effectively, leading to significant performance improvements.

Apps in ChatGPT: The Apps SDK

The new Apps SDK, available today in preview, allows developers to build and test fully interactive applications that run directly inside ChatGPT. This creates a powerful new surface for user interaction.

• Built on MCP: The SDK is built on the Meta-Call Protocol (MCP), which acts as a backbone to synchronize the app’s server, the model, and the UI.
• Bring Your Own Stack: Developers can use their existing web development skills and frameworks (like React) for the front end and any MCP-supported language for the back end.
• Interactive and Conversational: It enables a seamless, two-way communication channel. Users can interact with a rich visual UI, and the state of that UI is synchronized with the conversation. The model can pre-populate the app’s state based on the user’s request and update it based on subsequent conversation.
• Context-Aware: Apps inside ChatGPT can leverage the user’s context to provide personalized, relevant experiences, such as tailored recommendations.

Later this year, OpenAI will open submissions for developers to publish their apps, release a directory for discovery, and share more details on monetization.

• Tags:
• Developer Tools
• API
• GPT-5
• Codex
• Sora 2
• Apps SDK
• Agent Kit