OpenAI AgentKit Review 2026: Pricing, Features and Verdict

What it is

AgentKit is the name OpenAI gave in October 2025 to a bundle of tools for building agents on its API. The pitch at launch was a full stack: a drag and drop Agent Builder canvas, ChatKit for embedding chat, a Connector Registry, guardrails, and evals on top of the Agents SDK.

A year later the bundle looks different. On June 3, 2026 OpenAI told developers Agent Builder is deprecated, and its deprecations page lists November 30, 2026 as the shutdown date. The documented path forward is to export each workflow as Agents SDK code in Python or TypeScript, or to paste that export into ChatGPT and convert it into a Workspace Agent, which requires a ChatGPT Business, Enterprise or Edu workspace. OpenAI's own migration guide warns that the export does not convert the workflow graph and does not guarantee every behavior transfers unchanged.

What remains is a developer kit with four parts:

• Agents SDK. Open source libraries for Python and TypeScript that run the agent loop, tools, handoffs and guardrails inside your app.
• Agents API (beta). A managed service that runs OpenAI's Codex harness, with durable sessions, automatic context compaction, multi agent orchestration, MCP servers and sandboxes.
• ChatKit. An embeddable chat interface with widgets, file attachments, tool invocation and reasoning display.
• Hosted tools and evals. Web search, file search, hosted shell and Code Interpreter containers, plus trace grading to score agent runs.

Most of the top search results for this keyword were written at launch and still review Agent Builder as the centerpiece. That advice is out of date.

Pricing explained

There is no seat, plan or platform fee anywhere in AgentKit. Every cost is a meter on OpenAI's API pricing page.

Item	Price, October 2026
gpt-6-luna	$0.10 input, $0.50 output per 1M tokens
gpt-6.1-sol	$2.00 input, $10.00 output per 1M tokens
gpt-6-astra	$10.00 input, $50.00 output per 1M tokens
gpt-5.5	$5.00 input, $30.00 output per 1M tokens
Web search	$10 per 1,000 calls plus search content tokens
File search	$2.50 per 1,000 calls; storage $0.10 per GB per day after 1 GB free
Hosted shell and Code Interpreter	$0.03 (1 GB) to $1.92 (64 GB) per 20 minute container session
ChatKit uploads	$0.10 per GB day after 1 GB free per month
Agents SDK	Free, open source
Agents API (beta)	Model, tool and container usage only

Prices are standard rate, short context. Cached input is cheaper: $0.10 per 1M on gpt-6.1-sol, a 95% discount, which matters for agents that resend the same long system prompt every turn. Batch and flex processing cut rates roughly in half for work that does not need an instant answer.

Worked examples

Take a support agent that answers four turns per conversation, sending about 3,000 input tokens and getting 500 output tokens per turn, with one file search call per conversation.

• On gpt-6.1-sol: 12,000 input tokens cost $0.024, 2,000 output tokens cost $0.020, and the file search call costs $0.0025. That is about $0.047 per conversation, or about $465 for 10,000 conversations a month.
• On gpt-6-luna: the same conversation costs about $0.0012 in input, $0.001 in output and $0.0025 for file search, or about $0.0047. 10,000 conversations come to about $47.
• With caching on gpt-6.1-sol: if 2,500 of each 3,000 input tokens are a cached system prompt, input drops to about $0.0050 per conversation and the total to roughly $0.028.

For comparison, Agentforce lists $2 per conversation for customer facing agents, and Intercom Fin charges $0.99 per resolved outcome. Raw API pricing is far cheaper per conversation, but you pay in engineering time instead. See the AI Price Index for the category spread.

Hidden costs

• Engineering. Without Agent Builder, every new agent is a code project with hosting, logging and deployment.
• Runaway loops. An agent that calls web search ten times per answer adds $0.10 in tool fees alone. Set iteration limits.
• Storage. File search storage is billed per GB per day, so a 20 GB knowledge base costs $1.90 a day, about $57 a month.
• Migration. If you built on Agent Builder, budget time to re-implement deterministic branches that the export does not preserve.

Key features

Agents SDK

The Agents SDK is the center of AgentKit now. You define agents with instructions, tools and handoffs to other agents, and its runner handles the loop. It supports guardrails that run alongside an agent, sessions for memory, tracing to OpenAI's dashboard, and MCP servers as tools. The Python and TypeScript repositories are public on GitHub, and the docs cover running other model providers through the same interfaces.

Agents API

The Agents API, still in beta, is OpenAI's answer to managed agent runtimes like Amazon Bedrock AgentCore. OpenAI runs the Codex harness: you create a session, give it a task, stream events or receive a webhook, and steer it mid turn. It includes context compaction for long tasks, subagents, programmatic tool calling and a sandbox where the agent can run code, edit files and produce artifacts. You can use an OpenAI hosted sandbox, your own sandbox, or none.

ChatKit

ChatKit is a ready made chat front end. It handles streaming, file attachments, widgets for structured output and display of tool calls. For new apps OpenAI recommends the advanced integration, where ChatKit talks to your own server and agent code. The simpler hosted path, where ChatKit points at an Agent Builder workflow ID, only makes sense during the transition window.

Hosted tools

Agents can call OpenAI's web search, file search over vector stores, a hosted shell and Code Interpreter without you running that infrastructure. Each has a published per call or per session rate, listed above.

Evals and trace grading

Trace graders score full agent runs, not only single answers, so you can check whether an agent picked the right tool and followed policy. Prompt optimization and datasets sit in the same dashboard.

Safety guidance

OpenAI publishes guidance on prompt injection and data leakage for agent workflows, and the SDK supports input and output guardrails. These are tools, not guarantees; you still own the threat model for any agent that can act on private data.

What users say

We could not find a G2, Capterra or Trustpilot listing for AgentKit with 10 or more reviews, so there is no rating to show. Developer discussion since launch has a consistent pattern:

• Praise: fast prototyping, no platform fee, tight integration with OpenAI's models and tracing, and ChatKit saving weeks of front end work.
• Complaints: lock in to OpenAI's API for hosted pieces, a canvas that felt limited next to n8n or Make for real automation, and now frustration that Agent Builder is being retired a little over a year after launch.

Several launch reviews asked whether AgentKit would replace Zapier or n8n. The deprecation answers that: OpenAI has stepped back from the no code workflow market and pushed business users toward Workspace Agents in ChatGPT.

Privacy and security

• Training: OpenAI states that data sent to the API is not used to train its models unless you opt in, a policy in place since March 1, 2023.
• Retention: abuse monitoring logs, which can contain prompts and outputs, are kept for up to 30 days by default. Approved customers can get Modified Abuse Monitoring or Zero Data Retention. Under Zero Data Retention the store parameter on Responses is always treated as false.
• Application state: features like file search vector stores, ChatKit uploads and Agents API sessions store data to work. Delete them when done.
• Self hosting: you run the Agents SDK and a custom ChatKit server in your own infrastructure, but model calls still go to OpenAI unless you point the SDK at another provider.

Who it is for

• Developers and product teams already building on OpenAI models who want an agent framework with no extra license.
• Startups embedding a support or onboarding agent in their own app, where ChatKit saves front end work.
• Teams that want a managed runtime for long running agent tasks and are comfortable with a beta API.
• Companies with a ChatGPT Enterprise contract that want a code path alongside Workspace Agents.

Who should skip it

• Non technical teams. With Agent Builder gone, there is no no code path. Look at Zapier, Make or Lindy.
• Microsoft 365 shops. Copilot Studio gives low code agents with Entra, Purview and Teams built in.
• Salesforce or ServiceNow customers. Agents that need to act on CRM or ITSM records are easier in Agentforce or ServiceNow AI Agents.
• Teams that need model choice by default. n8n and Relevance AI let you swap models per step without writing an SDK integration.
• Anyone still planning to start on Agent Builder. Do not; it shuts down November 30, 2026.

Alternatives

• n8n: visual workflows with AI agent nodes, any model and free self hosting. See n8n pricing.
• Microsoft Copilot Studio: low code agents for Microsoft 365, billed in Copilot Credits.
• Gemini Enterprise: Google's employee agent platform with no code agents over 100+ connectors.
• Amazon Bedrock AgentCore: framework agnostic managed runtime on AWS, billed per vCPU hour.
• Relevance AI: no code agent teams for sales and operations.

For the full ranking, see best AI agent builders and best enterprise AI agent platforms.

Verdict

AgentKit is now a developer toolkit, not a builder. Judged that way it is strong: no platform fee, clear per token and per call prices, a managed agent runtime, a production chat UI and API data that is not used for training by default. The catch is that OpenAI retired the piece that made AgentKit approachable to non developers barely a year after launch, so weigh platform risk before you build anything that depends on a hosted OpenAI workflow. Use the Agents SDK if you want code you control; use Workspace Agents if you want no code inside ChatGPT; and pick a dedicated builder if neither fits.