OpenAI’s latest blog posts reveal significant advancements across model capabilities, developer tools, and workflow automation. Key trends and announcements include:

  • Enhanced Automation and Workflow Integration: New workspace agents (in research preview) enable users to automate repeatable, end-to-end business tasks using integrations with tools like Google Calendar and SharePoint. Multi-agent and multi-tool workflows are now further empowered by structured outputs—strict JSON schema enforcement—delivering more reliable and consistent automation, especially for complex, collaborative, or data-driven scenarios.

  • Expanded Model Capabilities: Several posts highlight improvements in model performance and specialization. Notably, GPT Image introduces advanced, photorealistic image generation and editing features. New fine-tuning options, including distillation (gpt-4o-mini) and reinforcement fine-tuning, facilitate domain-specific adaptation—such as for medical reasoning or faster, cost-effective classification tasks without sacrificing accuracy.

  • Strengthened Developer Experience: OpenAI has launched new APIs and SDKs (e.g., the Responses API, Agents SDK VoicePipeline) prioritizing usability, flexibility, and multi-modal support (text, image, audio). The Responses API, for instance, simplifies stateful, multi-turn interactions and integrates hosted tools like search out-of-the-box, streamlining workflow construction and reducing complexity.

  • Robust Infrastructure and Best Practices: Guidances on managing API rate limits, model output formatting, and optimizing tool performance (e.g., with the Model Context Protocol or eval-based iteration) help developers build scalable, reliable applications, aligning technical effectiveness with real-world business requirements.

  • Frontier Use Cases: Posts demonstrate rapid full-stack application development (aided by GPT-5) and advanced information retrieval (via Retrieval-Augmented Generation and Pinecone integration), underscoring OpenAI’s push toward enabling dynamic, context-aware, and production-ready AI systems.

Collectively, these developments signal a push towards highly integrative, customizable, and accessible AI solutions—equipping developers and enterprises to build sophisticated, automated, multi-modal applications at scale.

New Cookbook Recipes

chatgpt-agents-sales-meeting-prep.md

Source: openai/openai-cookbook

The blog post introduces the development of ChatGPT workspace agents, designed to facilitate repeatable end-to-end tasks, currently in research preview for Business, Enterprise, and Edu customers. These agents enable users to automate workflows by accessing connected applications like Google Calendar and Microsoft SharePoint. The guide outlines how to build a workspace agent to assist sales teams by automatically generating meeting briefs from customer data and company news. It covers prerequisites, the agent-building process, including workflow configuration, app permissions, and skills for consistent outputs. The testing, scheduling, and sharing of agents are also addressed, allowing for automated daily updates and collaborative use among team members. These agents enhance productivity by streamlining preparation for sales meetings.

Generate_Images_With_GPT_Image.ipynb

Source: openai/openai-cookbook

The blog post introduces GPT Image, a new large language model with advanced image generation capabilities and improved photorealism compared to previous models like DallE 2 and 3. Users can generate images from detailed prompts, edit images using existing inputs, and customize output properties such as quality, size, and background transparency. The model allows for intricate image creation, utilizing capabilities like instruction-following and the generation of masks for specific edits. Additionally, the post outlines setup instructions and provides code examples for generating images, applying masks, and editing based on user specifications. A link to an image gallery for inspiration is also provided, encouraging users to explore further use cases.

How_to_handle_rate_limits.ipynb

Source: openai/openai-cookbook

The blog post discusses strategies for managing rate limits when using the OpenAI API, which helps prevent overload and ensures fair access for all users. Key features highlighted include the default automation of rate limits based on usage, and guidance on handling errors like 429: 'Too Many Requests'. Effective mitigation strategies include implementing exponential backoff for retries, adjusting max_tokens to better align with expected responses, and employing batching techniques for requests to improve throughput. The post also introduces libraries such as Tenacity and Backoff for ease of implementation, alongside a link to an example script for parallel processing API requests. Users are encouraged to thoroughly test fallback models and batching methods to ensure they meet application requirements while adhering to rate limits.

Leveraging_model_distillation_to_fine-tune_a_model.ipynb

Source: openai/openai-cookbook

The blog post discusses the process of model distillation, specifically using OpenAI’s outputs from gpt-4o to fine-tune a smaller model, gpt-4o-mini, aimed at reducing latency and costs for specific tasks. The author demonstrates this through a classification task involving French wine reviews, where the model predicts grape varieties based on provided attributes.

Key actions include analyzing a dataset, generating prompts, and employing structured outputs for constrained responses. Performance comparisons show that gpt-4o significantly outperforms gpt-4o-mini. After distillation, the fine-tuned model achieves a notable accuracy improvement, offering about a 22% relative enhancement over the non-distilled gpt-4o-mini, thus ensuring efficiency without compromising accuracy.

Reinforcement_Fine_Tuning.ipynb

Source: openai/openai-cookbook

The blog post outlines a guide for developers utilizing OpenAI’s reinforcement fine-tuning (RFT) techniques to enhance the reasoning capabilities of their models. Focused specifically on the o4-mini reasoning model, the post details a case study in life sciences, where the aim is to improve the model’s accuracy in predicting outcomes from doctor-patient transcripts.

Key sections of the guide include setting up the environment, gathering appropriate datasets, benchmarking model performance, defining grading systems, and training processes. Notably, RFT is presented as a valuable approach for situations requiring high precision, such as medical data interpretation, leveraging small datasets effectively. The guide emphasizes the importance of effective grading mechanisms to guide learning and avoid common pitfalls. Ultimately, it serves as a comprehensive manual for practitioners looking to apply RFT in specific domains.

Speech_transcription_methods.ipynb

Source: openai/openai-cookbook

This blog post presents a practical guide for beginners to utilize Speech-to-Text (STT) capabilities with the OpenAI API, highlighting various methods along with their use cases, advantages, and limitations. Key methods include:

  1. File Upload (Blocking): Best for voicemail and meeting recordings, but lacks partial results.
  2. File Upload (Streaming): Offers real-time updates, ideal for voice memos, yet requires complete audio files upfront.
  3. Realtime WebSocket: Enables live captioning with ultra-low latency, though it demands audio format compliance and session management.
  4. Agents SDK VoicePipeline: Facilitates real-time transcription and integration into workflows with minimal setup, currently in Python-only beta.

The blog emphasizes choosing the right STT method based on use-case needs, trade-offs in latency, and the evolving nature of the API features. Users are encouraged to experiment and provide feedback.

Structured_Outputs_Intro.ipynb

Source: openai/openai-cookbook

The blog post introduces the Structured Outputs feature in the Chat Completions API and Assistants API, enabling models to generate responses that strictly adhere to specified JSON schemas by setting the strict: true parameter. Key enhancements include the ability to define strict response formats and function schemas, ensuring output consistency and adherence to complex workflows. The article provides examples demonstrating the practical applications of Structured Outputs, such as delivering step-by-step solutions for math tutoring, summarizing articles, and extracting user input for product recommendations. The post emphasizes the robustness this feature adds to production-level applications, maintaining output integrity and usability across various scenarios. Structured Outputs will only be available with specific models, including gpt-4o-mini and gpt-4o-2024-08-06.

Structured_outputs_multi_agent.ipynb

Source: openai/openai-cookbook

The blog post introduces Structured Outputs for building multi-agent systems, enhancing JSON mode with a strict schema enforcement using the strict: true parameter. This capability is essential when managing multiple tools, as it improves performance by allowing specialized agents to handle specific tasks.

The post outlines a multi-agent architecture consisting of four agents: a Triaging Agent, a Data Pre-processing Agent, a Data Analysis Agent, and a Data Visualization Agent. It details how these agents will cooperate to perform data analysis tasks, including cleaning data, statistical analysis, and generating visualizations like line charts.

Each agent is equipped with specific tools aligned with its role, and example code illustrates the execution logic necessary for interacting with user queries and processing data. The post emphasizes the importance of structured outputs in enhancing the efficiency of complex, multi-agent systems.

gpt-5_frontend.ipynb

Source: openai/openai-cookbook

The blog post discusses the advancements in frontend development facilitated by GPT-5, highlighting its ability to create full-stack applications rapidly and perform complex code edits effortlessly. It provides a prompt guide for effective usage and recommends various frameworks and libraries, including Next.js, React, and Tailwind CSS, to enhance GPT-5’s capabilities.

An interactive example demonstrates how to create frontend applications from scratch using simple prompts, showcasing the model’s steerability and multimodal input processing. The ability to utilize both text and images improves design consistency when generating web pages. Additionally, the post underscores the potential for generating dynamic content, like games, while emphasizing the necessity for user review and iteration to achieve production-quality results. The author encourages readers to experiment with the tools and share their creations.

mcp_tool_guide.ipynb

Source: openai/openai-cookbook

The blog post discusses the hosted Model Context Protocol (MCP) tool integrated within the Responses API, which simplifies the interaction between applications and external services. Instead of traditional function calling—which incurs latency and complex management—the MCP tool allows for direct interaction with an MCP server that centralizes commands like “search product catalog” or “add item to cart,” reducing network hops.

Key highlights include several use cases that benefit from MCP, such as streamlined commerce transactions and effective error reporting in DevOps. It outlines the operational steps for using the MCP tool, emphasizes best practices for optimizing performance (like filtering tools to reduce token payloads), and illustrates the tool’s integration with other systems through a comprehensive pricing analysis scenario. The MCP tool serves as a robust solution to enhance efficiency when building applications that require external service interactions.

receipt_inspection.ipynb

Source: openai/openai-cookbook

The blog post presents a comprehensive guide for using evals in designing autonomous systems, particularly focusing on replacing labor-intensive workflows, such as receipt analysis for expense validation. It targets ML/AI engineers and Solution Architects, offering a modular and executable notebook that outlines a step-by-step methodology, from starting with a small labeled dataset and building an MVP to leveraging evals for continuous iteration and improvement. The narrative emphasizes business alignment through cost-benefit analyses of eval performance, ensuring that development efforts are directed toward high-impact improvements. Key components include understanding the problem, gathering examples, building initial systems, mapping evals to business metrics, and integrating QA processes. The post also details a use case involving receipt parsing, highlighting the project lifecycle and the importance of ongoing evaluation to maintain system performance.

responses_api_tool_orchestration.ipynb

Source: openai/openai-cookbook

The blog post presents a comprehensive guide for creating dynamic multi-tool workflows using OpenAI’s Responses API, emphasizing the Retrieval-Augmented Generation (RAG) approach. Key features include intelligent routing of user queries to suitable tools, whether for general knowledge or specific context retrieval from external databases like Pinecone.

A practical example of using the Responses API’s file search feature with PDF documents is provided, showcasing its flexibility. The post details how to process datasets to generate embeddings for querying a Pinecone index, upsert data, and execute semantic searches. Additionally, it discusses the orchestration of multiple tool calls, including a web search tool and a Pinecone document search tool. This approach ensures contextually accurate, data-driven responses to user queries, facilitating complex information retrieval tasks effectively.

responses_example.ipynb

Source: openai/openai-cookbook

The blog post introduces the Responses API, a newly developed interface for interacting with OpenAI models that prioritizes simplicity and flexibility. Key features include:

  • Seamless Multi-Turn Interaction: Enables continuous conversation with state management handled by the API, allowing developers to focus on other aspects of their applications.
  • Hosted Tools: Integrates various tools like file search and web search directly, thereby enhancing functionality without additional setup.
  • Multimodal Support: Capable of processing and responding to text, images, and audio inputs, enabling more complex interactions within a single API call.

Practical examples demonstrate how to initiate conversations, manage contexts, and perform advanced tasks, such as web searches related to images—significantly streamlining workflows compared to previous APIs. This positions the Responses API as a robust solution for developers seeking to create sophisticated AI applications.