Overview

OpenAI introduced the new Responses API to unify and simplify complex agentic workflows, while maintaining the Chat Completions API and sunsetting the Assistance API by 2026, giving developers a full year to transition.

Three powerful built-in tools were announced: a web search tool with 90% accuracy (up from 38%), an improved file search tool with expanded file support and metadata filtering, and a computer use tool enabling AI to interact with screens.

The web search capability allows real-time data structuring from web sources with citations, costing approximately $30 per 1,000 queries, with various strategies available for managing search costs.

OpenAI's Agents SDK now supports types, guard railing, and tracing capabilities viewable in the dashboard, enabling more modular agent design with easier monitoring and visualization of workflow steps.

With these developments, OpenAI is positioning 2025 as "the year of agents" with a strategy to gradually merge successful preview models into the main branch, similar to how vision capabilities were integrated into GPT-4o.

Content

OpenAI API Launch and New Tools

OpenAI announced three new built-in tools:

* Web search tool (similar to ChatGPT for search) * Improved file search tool * Computer use tool (from ChatGPT Operator product)

Introducing the new Responses API:

* Designed to support more complex, multi-turn agentic workflows * Aims to unify functionality from chat completions and assistance API * Simplifies tool integration for developers * Will support everything chat completions and assistance APIs currently support * Offers a stateless mode (by passing store=false) * Stores conversation state for free for 30 days * Provides visual debugging and observability in dashboard

Important API transition details:

* Chat completions API is NOT going away - Will continue to be maintained - Optimized for earlier, simpler text-based interactions * Assistance API has a planned sunset date in first half of 2026 * OpenAI will provide: - Smooth migration path - Full year for developers to transition - Additional features like assistant objects and thread-like objects - Future additions of code interpreter tool and async mode

API recommendations:

* New users should start with Responses API * Offers more capabilities and performance than chat completions * Chat completions will still be supported for years

Web Search Feature Details

Launched in two ways:

* As a tool in Responses API * Direct access to fine-tuned search model (GPT-4o Search Preview) in chat completions

Performance improvements:

* Accuracy increased from 38% to 90% in simple QA * Search team focused on: - Gathering information from multiple data sources - Selecting and citing information accurately - Using synthetic data and model distillation techniques

Web Search capabilities:

* Can be combined with other tools like function calling and structured outputs * Allows for real-time data structuring from web sources * Provides citations from web sources * Comparable to similar APIs from Perplexity and Gemini

Technical considerations:

* Knowledge cutoff varies depending on use case * Currently no built-in parameter for search depth/breadth * Potential for agent orchestration to explore deeper search layers * Cost is approximately $30 per 1,000 queries * Potential strategies for managing search costs include: - Context budget approach - Similarity matching cut-offs - Storing search results in files to avoid repeated searches

Emerging use cases:

* Companies like Hebea using web search for accessing public information * Potential for storing user preferences/memories in vector stores

File Search Enhancements

File search capabilities:

* Can be used to find personalized recommendations based on user preferences * Combining with neural networks and real-time internet access enables precise, context-aware answers * Allows integration of private company documents with AI systems * Can be combined with web search for more dynamic information retrieval

New features:

* Expanding file type support * Query optimization * Custom re-ranking * Metadata filtering becoming available (critical for large vector stores)

Implementation approaches:

* OpenAI offers an out-of-the-box file search solution with some customization options * Recommendation: Start with managed solution, then customize or switch to custom solution if needed * Some AI engineers prefer building their own vector database stack for more control

Computer Use/Operator Tool

Developed a custom model optimized for computer use
Currently in early stages (compared to GPT-1/2 level of development)
API enables agents to:

* Interact with screen (click, scroll, type) * Complete multi-step tasks * Report back on actions * Uses screenshot inputs to determine actions

Potential applications include task automation for products and customers
Discussed Pokemon as a potential agent benchmark

Model Development Strategy

Fine-tuned models initially operate separately
Goal is to merge successful preview models into main branch over time
Similar to how vision capabilities were integrated into GPT-4.0
Aim to reduce model fragmentation

Agents SDK Updates

Added support for:

* Types * Guard railing (parallel execution with blocking capability) * Tracing (viewable in OpenAI dashboard)

Flexible design allows:

* Integration with any chat completions API provider * Multiple tracing provider support

Development background:

* Originated from customer feedback about agent orchestration challenges * Enables more modular agent design with easier monitoring * Supports creating triage agents that route to different specialized agents

Troubleshooting and Improving Agentic Workflows

Key capabilities include:

* Ability to trace workflow steps between agents * Visualizing tool calls and handoffs between different agents * Helping developers build more effective AI agent systems

Future plans involve:

* Connecting traces to evaluation (evals) products * Using traces to generate better evaluations * Potentially implementing reinforcement fine-tuning based on those evaluations

The overall goal is to create tools that make agent development easier and more transparent for developers

OpenAI is positioning 2025 as the "year of agents" with these new tools and APIs designed to support more complex AI agent development

⚡️The new OpenAI Agents Platform