I Compared All the OpenAI APIs — Responses API, Chat Completions, Agents SDK, AgentKit: Which One Should You Actually Use?
This page has been translated by machine translation. View original
Introduction
Have you noticed how much OpenAI's APIs have grown?
When I started researching how to build an agent that uses web search and Code Interpreter, I ran into so many options — Responses API, Chat Completions API, Assistants API, Agents SDK, AgentKit… — that I found myself completely lost on which one to use.
In this article, I'll organize the overall landscape of OpenAI APIs as of June 2026, and summarize the criteria for deciding "which one should I actually choose."
Laying Out All of OpenAI's APIs
First, to get a clear picture of the overall landscape, here's a list of the major APIs available as of June 2026.

| API | Status | Main Use Case |
|---|---|---|
| Responses API | Recommended (Default) | Built-in tools (web search, Code Interpreter, file search), agent loops, MCP, Deep Research, Computer Use |
| Chat Completions API | Supported | Simple request/response. No built-in tools, conversation state managed manually |
| Assistants API | Deprecated (Shutdown 2026/8/26) | Stateful assistants — replaced by Responses API |
| Realtime API | Active | Low-latency voice agents (speech-to-speech), SIP phone, image input |
| Batch API | Active | Asynchronous bulk processing, 50% cost reduction, webhook-driven |
| Legacy Completions API | Discontinued | Old /v1/completions endpoint |
And as frameworks/toolkits built on top of the APIs:
| Tool | Status | Role |
|---|---|---|
| Agents SDK | Active (open source) | Python/TS library. Agent orchestration |
| AgentKit | Partially sunset planned | Visual toolkit (Agent Builder + ChatKit + Connector Registry) |
What surprised me was that there are 6 types of APIs alone, plus 2 framework layers on top. In practice, though, when you narrow it down to "what should I consider first for a new project," it becomes much simpler.
Responses API vs Chat Completions API — What's the Difference?
This is the comparison that trips up the most developers. After looking into it, I found the differences are much bigger than I expected.
| Aspect | Responses API | Chat Completions API |
|---|---|---|
| State management | Can be managed server-side | Must send conversation history yourself |
| Built-in tools | Web search, Code Interpreter, file search | None |
| MCP support | Yes | No |
| Deep Research | Yes | No |
| Computer Use | Yes | No |
| Cost | 40-80% improved cache efficiency | Baseline |
| Reasoning model performance | +3% on SWE-bench (when using GPT-5) | Baseline |
Particularly noteworthy is the cost difference. Because the Responses API can hold conversation state server-side, the cache hit rate improves dramatically. Internal tests have reported a 40–80% improvement in cache efficiency.
When to Use the Chat Completions API
There are still cases where the Chat Completions API is the right choice:
- Simple one-off request/response
- When there's no immediate need to migrate existing Chat Completions integrations
- When compatibility with third-party LLM providers is a priority (the Chat Completions format is close to an industry standard)
However, both OpenAI and Microsoft recommend the Responses API by default for new projects.
Why Is the Assistants API Going Away?
The Assistants API launched in 2024 and was expected to be the go-to API for building stateful assistants. However, because the Responses API absorbed its functionality as a superset, it will be shut down on August 26, 2026.
Key migration points:
- Conversation management via Assistants API "Threads" → equivalent functionality with the
previous_response_idparameter in the Responses API - Assistants API "Tools" → integrated into Responses API built-in tools
- Assistants API "Files" → integrated into Responses API file search tool
For projects already using the Assistants API, refer to the OpenAI official migration guide and plan your migration accordingly.
Agents SDK vs AgentKit — Code-First vs Low-Code
When building agents, another choice that can cause confusion is between the Agents SDK and AgentKit.
| Aspect | Agents SDK | AgentKit |
|---|---|---|
| Approach | Code-first | Low-code / drag & drop |
| Key components | Agents, Handoffs, Guardrails, Tools, Sessions | Visual Canvas, Connector Registry, ChatKit, Evals |
| Multi-agent | In-code delegation via Handoffs | Visual node configuration |
| External integrations | Custom implementation via tools/MCP | Connector Registry (Slack, Google Drive, etc.) built-in |
| UI components | Build your own | Provided by ChatKit |
| Provider lock-in | None (can use other models too) | OpenAI ecosystem only |
| Future outlook | Actively in development | Agent Builder & Evals sunset 2026/11/30 |
Important Note: Partial Sunset of AgentKit
On June 3, 2026, OpenAI announced the end of the Agent Builder and Evals products. They will no longer be available on the OpenAI platform after November 30, 2026.
This is a significant factor in the decision. Workflows that depend on AgentKit's visual builder will need to be migrated to another solution within six months.
Decision Criteria
When to choose Agents SDK:
- You want to control agent logic in code
- You need multi-agent Handoffs, typed tools, and schema validation
- You want to use models other than OpenAI (provider-agnostic)
- You need to handle long-running tasks with Sessions
- You want to manage "workflow = code" in a production system
When to choose AgentKit:
- Quick prototyping
- A non-engineer team wants to build simple agents
- You want to quickly implement a built-in chat UI with ChatKit
- You want to easily configure integrations with existing SaaS tools via the Connector Registry
Honestly, given the partial sunset announcement for AgentKit, Agents SDK is the safer long-term choice. AgentKit is convenient for prototyping and short-term projects, but relying on it in production carries risk.
Summary: A Selection Guide for June 2026
After looking into everything, the decision can be summarized simply:
| What you want to do | API/Tool to choose |
|---|---|
| New projects in general | Responses API |
| Simple one-off requests | Chat Completions API (but Responses is recommended for new projects) |
| Voice agents | Realtime API |
| Large-scale batch processing | Batch API |
| Building agents (code) | Agents SDK |
| Building agents (low-code) | AgentKit (watch out for sunset) |
| Existing Assistants API | Migrate to Responses API by 2026/8/26 |
OpenAI's API strategy is converging around the "Responses API as the center." The Chat Completions API will continue to be supported, but new features such as built-in tools, MCP, and state management will only be available in the Responses API.
For new projects, the combination of Responses API + Agents SDK is the most future-proof choice available right now.
