OpenAI AgentKit: What It Is and What Just Changed

OpenAI AgentKit is OpenAI's toolkit for building and shipping AI agents, launched on October 6, 2025, and it is in the middle of a major shake-up: on June 3, 2026, OpenAI announced it is winding down Agent Builder and the Evals platform, with both shutting down on November 30, 2026, while ChatKit remains available. If you are evaluating AgentKit today, the practical picture is a deprecated visual builder, a surviving embeddable chat UI, and a recommended migration to the code-first OpenAI Agents SDK or ChatGPT Workspace Agents. This guide covers what AgentKit includes, what it costs, exactly what is being retired, and how the surviving pieces compare to Anthropic's Claude Agent SDK.

What Is OpenAI AgentKit?

OpenAI AgentKit is a set of building blocks for creating, deploying, and optimizing AI agents on the OpenAI platform, announced by Sam Altman at OpenAI's DevDay on October 6, 2025. Altman described it as "a complete set of building blocks" designed to help developers "take agents from prototype to production with way less friction," per TechCrunch's launch coverage.

The launch pitch was speed: an OpenAI engineer built a working two-agent workflow live onstage in under eight minutes using Agent Builder's drag-and-drop canvas, which OpenAI compared to "Canva for building agents." Under the hood, AgentKit workflows run on OpenAI's Responses API, the same API surface that powers tool use and reasoning across OpenAI's model lineup, as documented in the official AgentKit announcement.

Nine months later, the story changed. As of July 2026, AgentKit is best understood as two products that survived (ChatKit and the Connector Registry) and two that did not (Agent Builder and the hosted Evals platform).

What Is Inside AgentKit

AgentKit shipped with four components, and as of July 2026 exactly half of them are being retired. Here is each piece, what it does, and its current status.

The component people usually mean when they search for "OpenAI AgentKit" is Agent Builder, the visual workflow canvas. That is the piece being removed, which is why the deprecation matters more than a typical product sunset.

The June 2026 Deprecation, Explained

OpenAI announced on June 3, 2026 that Agent Builder and the Evals platform are being deprecated, with both shutting down on November 30, 2026, per the official OpenAI deprecations page. The Evals platform goes read-only earlier, on October 31, 2026, which is the effective deadline for exporting evaluation datasets and graders.

OpenAI's recommended migration paths are:

• For workflows that should live as code: the OpenAI Agents SDK, the open source Python and TypeScript framework built on the Responses API.
• For natural-language, non-developer use cases: Workspace Agents inside ChatGPT.
• For evals: OpenAI's migration guidance references moving to Promptfoo, an open source evaluation framework, or building custom evaluation scripts.

Three practical notes if you have agents running in Agent Builder today. First, existing agents keep working through the migration window, so you can run the old and migrated versions side by side before cutting over. Second, export your workflows, prompts, and configurations before November 30, 2026, because there is no visual replacement. Third, tool integrations built on MCP servers continue working unchanged, which is a strong argument for keeping integrations on open protocols rather than proprietary surfaces.

ChatKit is explicitly unaffected: you can keep shipping a ChatKit frontend, but once the hosted Agent Builder backend is gone, it needs a backend you control.

OpenAI AgentKit Pricing

OpenAI AgentKit has no separate platform fee: you pay standard OpenAI API token rates for the model calls your agents make, and the AgentKit tooling itself is included with API access. That pricing model was true at launch and remains true for the surviving components as of July 2026.

The cost drivers are therefore the same as any agent workload: which model you route to, how many tool-call round trips your workflows make, and how much context each step carries. The Connector Registry adds an eligibility gate rather than a price: it began rolling out to API, ChatGPT Enterprise, and ChatGPT Edu customers. For a wider look at what agent stacks cost to run, see our AI agent pricing comparison.

AgentKit vs Claude Agent SDK

The core difference is that the Claude Agent SDK ships a complete agent harness (built-in file tools, bash execution, web search, permissions, MCP support, and the agent loop), while the surviving AgentKit stack gives you orchestration primitives via the Agents SDK plus a polished chat frontend via ChatKit, and you assemble more of the runtime yourself. We cover the Anthropic side in depth in our Claude SDK guide.

Both vendors have converged on the same posture: code-first SDKs, standard token pricing, and MCP as the integration layer. The remaining differences are the harness depth on Anthropic's side and the frontend kit on OpenAI's side. For the broader platform rivalry, see our Anthropic vs OpenAI comparison.

When to Use Which

Choose based on where your agent's work happens: pick the Claude Agent SDK when the agent operates on files, terminals, and codebases, and pick the OpenAI Agents SDK plus ChatKit when you are building a customer-facing conversational product on OpenAI models.

• Use the OpenAI Agents SDK + ChatKit if you are committed to OpenAI models, want a production chat UI quickly, and are comfortable expressing orchestration as code. This is also the mandatory path for existing Agent Builder users before November 30, 2026.
• Use the Claude Agent SDK if your agent needs to read, edit, and execute against a real environment out of the box, or if you want the same harness that powers Claude Code, including hooks, subagents, and CLAUDE.md project memory. Our Claude SDK guide has a working quickstart.
• Use neither directly if you want multi-agent coordination on top of Claude Code itself; that is what Claude Code agent teams and frameworks in our agent frameworks comparison cover.

Limitations and Tradeoffs

The obvious risk with AgentKit is platform stability: a flagship DevDay product lost half its surface area within nine months, and teams that built on the visual canvas now owe a migration by November 30, 2026. If your organization needs multi-year stability guarantees, weight open source frameworks and open protocols (like MCP) more heavily than hosted builders from any vendor. This guide is based on OpenAI's official announcement, deprecation documentation, and launch press coverage as of July 2026, not on hands-on production benchmarking of AgentKit workloads, and dates or migration guidance may change; check the OpenAI deprecations page before committing.

Related Guides

• Claude SDK Guide: Agent SDK vs Client SDKs Explained
• Anthropic vs OpenAI: The 2026 Platform Comparison
• AI Agent Frameworks Compared
• Everything Claude Code: The Complete Guide