Search for the OpenAI agent builder and you land in the middle of a product transition. The visual tool literally named Agent Builder is being retired on November 30, 2026, and the durable way to build agents on OpenAI is now code. In practice that means the open-source Agents SDK for orchestration, the Responses API with built-in tools for the runs themselves, and ChatKit for the chat interface. This guide explains each piece, what the metered pricing works out to, who the stack genuinely fits, and what to use instead if you want a working agent rather than a framework for building one. Every OpenAI fact here is verified from OpenAI's own documentation as of August 6, 2026. If agents themselves are new territory, start with our primer on what an AI agent is.

What does "OpenAI agent builder" actually mean?

The phrase covers three different products, and knowing which one you mean saves hours. It can refer to Agent Builder, a specific visual canvas now scheduled for shutdown; to OpenAI's broader developer stack for building agents; or to agent mode inside ChatGPT, the consumer feature. Most people searching mean the developer stack.

The OpenAI agent stack in 2026, piece by piece

Four names carry the load, and their statuses differ sharply as of August 2026. The Agents SDK and the Responses API are the active foundation, ChatKit is the interface layer, and Agent Builder, the visual canvas most people picture, is on a countdown to shutdown. Every fact below comes from OpenAI's own documentation.

PieceWhat it isShippedStatus, August 2026
Agents SDKOpen-source orchestration framework (Python and TypeScript, MIT license): agents, handoffs, guardrails, sessions, tracingMarch 11, 2025Active; OpenAI's recommended build path
Responses APIThe API primitive for agent runs, with built-in tools: web search, file search, computer use, remote MCP, shell, image generationMarch 11, 2025Active; gained multi-agent orchestration on July 9, 2026
ChatKitEmbeddable chat interface for agent apps: widgets, tool invocation, attachmentsOctober 6, 2025Active; named migration path for Agent Builder users
Agent BuilderVisual canvas for building multi-step agent workflows without codeOctober 6, 2025Deprecated June 3, 2026; shuts down November 30, 2026

Two more dates complete the picture. The Assistants API, the older agent-ish endpoint many tutorials still reference, shuts down on August 26, 2026, replaced by the Responses and Conversations APIs. And AgentKit, the DevDay 2025 umbrella brand, survives mainly through ChatKit and evaluation tooling now that its visual centerpiece is retiring; OpenAI's deprecation page points Agent Builder users to the Agents SDK or ChatGPT Workspace Agents instead.

This churn is strategy, not accident: OpenAI is consolidating everything onto the Responses API as the single agent primitive. We unpack the competitive logic in OpenAI vs Anthropic: agent strategy in 2026, and what the company's coding agents signal about where this is heading in what OpenAI's Codex agent means for the rest of us.

What you can build with it

The stack is genuinely capable when engineers drive it. The Agents SDK handles orchestration: which agent runs, when to hand off, what gets validated. The Responses API supplies the run itself plus hosted tools, so common capabilities arrive without infrastructure work on your side. Typical builds cluster into five shapes.

Connectivity is the other half of the pitch: remote MCP support means an agent can call third-party tool servers without you writing glue for each one. The ceiling here is high; the floor, as the next two sections show, is that someone on your team has to build and keep building.

What it costs

OpenAI charges metered API usage rather than a platform subscription: as of August 2026 the pricing page lists model token rates and per-tool charges, and no standalone fee for the agent tooling appears on it. Your bill is tokens consumed plus tool calls made, which makes volume the variable that matters.

Line itemPrice (August 2026)Notes
GPT-5.6-terra tokens$2.00 in / $12.00 out per 1MMid-tier model of the current GPT-5.6 family
GPT-5.6-luna tokens$0.20 in / $1.20 out per 1MSmall model for high-volume steps
GPT-5 tokens$1.25 in / $10.00 out per 1MPrevious-generation flagship, still listed
Web search tool$10.00 per 1,000 callsPlus retrieved content billed as tokens at model rates
File search tool$2.50 per 1,000 callsStorage $0.10 per GB per day, first 1 GB free
ChatKit upload storage$0.10 per GB-day after 1 GB freeThe only AgentKit-specific line on the pricing page
Hosted shell and code containersBilled by container size and durationSmall containers cost cents per session; larger sizes scale up

The bigger line item never appears on the pricing page: engineering time. A production agent needs design, prompt and tool iteration, evaluation suites, guardrails, tracing, and monitoring, and then maintenance through platform churn. Teams that priced only the tokens routinely find the salaries were the actual cost. That trade is worth making when the agent is your product; it is a poor trade when the agent is just a means to a finished task.

Who it fits

The stack fits teams for whom the agent is the product, or a core feature of it, and who have the engineers to own an agent loop in production. If that describes you, it is one of the strongest foundations available. Four profiles get the most from it.

Note the common thread: every profile assumes developers. That assumption is what separates a framework from a platform, a distinction we map fully in what is an AI agent platform.

Where it is not the right tool

Being a strong developer framework is exactly why it is the wrong tool for operators. If you want the outcome of an agent, the report compiled, the inbox triaged, without designing, coding, evaluating, and maintaining one, primitives are the wrong purchase. Four signals say build elsewhere, or do not build at all.

None of this is a criticism of the stack; it is a statement of who it is for. OpenAI sells excellent primitives for people who build. People who operate need the layer above.

Honest alternatives

The honest comparison set depends on what you are optimizing for: control, determinism, or the finished result. Developer frameworks compete on control, workflow tools on determinism, and agent platforms on outcomes. Price and skill requirements move accordingly, so match the category to whoever on your team will actually own the thing.

The two worlds also connect rather than compete. If you build agents with SDKs for a living, distribution is usually the harder problem than construction, and marketplaces exist to solve it: builders who publish agents on Gravity earn 21.27 percent of every run, rising to 25 percent at the top tier. Details are on the builders page, and the wider economics are in our guide to how to monetize AI agents.

Frequently asked questions

Is OpenAI's Agent Builder shutting down?

Yes. OpenAI announced the deprecation of Agent Builder on June 3, 2026, and its documentation states the product is scheduled to shut down on November 30, 2026. OpenAI points existing users to the Agents SDK or ChatGPT Workspace Agents as replacements, and ChatKit remains available.

Is the OpenAI Agents SDK free?

The SDK itself is open source under the MIT license, available in Python and TypeScript, so there is no license fee. You pay for what it drives: model tokens and built-in tool calls, billed at OpenAI's metered API rates. The framework can also route to other model providers.

What is the difference between AgentKit and the Agents SDK?

AgentKit is the umbrella OpenAI announced at DevDay on October 6, 2025, covering the visual Agent Builder, the embeddable ChatKit interface, and evaluation tooling. The Agents SDK is the open-source code framework released earlier, in March 2025, and it is the piece OpenAI now recommends for building agents as Agent Builder retires.

How much does it cost to build an agent on OpenAI?

As of August 2026 there is no platform subscription on OpenAI's pricing page; you pay metered usage. Examples: GPT-5.6-terra costs 2 dollars per million input tokens and 12 dollars per million output tokens, web search costs 10 dollars per 1,000 calls, and file search costs 2.50 dollars per 1,000 calls plus storage. Volume decides the bill, and engineering time is the larger hidden cost.

Do I need to code to use OpenAI's agent tools?

Increasingly, yes. The no-code visual canvas, Agent Builder, shuts down on November 30, 2026, and the recommended paths, the Agents SDK and the Responses API, are code-first in Python or TypeScript. Non-developers who want agent outcomes are better served by an agent platform with ready-made agents; our primer on what an AI agent is explains the distinction from first principles.

Is ChatGPT's agent mode the same as the developer platform?

No. Agent mode is a consumer feature inside ChatGPT that performs tasks for the person chatting. The developer platform, meaning the Agents SDK, the Responses API, and ChatKit, is for building your own agent applications on OpenAI's API and is billed separately as metered usage.

Sources