Join our Newsletter — 33% off our NHI Course
Home› Guides› OpenAI Agents SDK and AgentKit: Features, Versions, Pricing…
Guide AI Platforms

OpenAI Agents SDK and AgentKit: Features, Versions, Pricing and Deprecations

PDF for newsletter subscribers · free to read online

← All guides
By NHI Mgmt Group, edited by Lalit Choda Updated 15 November 2026 15 min read
On this page

The OpenAI Agents SDK is an open source Python and TypeScript library from OpenAI for building agents and multi-agent workflows, and AgentKit is the wider set of OpenAI agent tools announced on 6 October 2025. The two names are often used as if they were one product, and the picture changed in 2026: OpenAI's deprecations page now lists Agent Builder, the visual canvas inside AgentKit, as deprecated with a shutdown date of 30 November 2026, while the Agents SDK and ChatKit remain available. This page sets out what each part does, which versions are current, how the SDK is versioned, what it costs to run, how tracing handles data and what has changed since launch. It is written for developers, architects and the people who approve frameworks for a team. Every fact comes from OpenAI's own documentation, repositories, announcement and pricing page. Last checked: 2026-11-15.

This is an independent guide from NHI Mgmt Group, based on OpenAI's public documentation as of November 2026. It is not sponsored or reviewed by OpenAI. Product features change, so confirm settings against OpenAI's current documentation.

Key takeaways

  • The SDK is free and MIT licensed. Both openai-agents (Python) and @openai/agents (TypeScript) carry the MIT licence. You pay only for model tokens and any hosted tools the agent calls.
  • Agent Builder is being retired. OpenAI's deprecations page gives a notice date of 3 June 2026 and a shutdown date of 30 November 2026, and names the Agents SDK or ChatGPT Workspace Agents as replacements. It also states that "ChatKit remains available."
  • Still on 0.x, with breaking changes in minor versions. The release policy uses the form 0.Y.Z: a new Y can break public interfaces not marked beta, so pin minor versions and read the breaking change log before upgrading.
  • Tracing is on by default. Traces, including model and tool inputs and outputs, go to OpenAI unless you disable tracing or set trace_include_sensitive_data to false. Tracing is unavailable to organisations under a Zero Data Retention policy.
  • Not tied to OpenAI models. The SDK uses the Responses API by default but also supports Chat Completions compatible endpoints and beta adapters for LiteLLM and Any-LLM.

What is OpenAI Agents SDK and AgentKit?

The documentation describes the Agents SDK as a lightweight framework for building agentic apps in an "easy-to-use package with very few abstractions", and as a "production-ready upgrade" of Swarm, OpenAI's earlier experimental agent project. The repository README calls it "provider-agnostic, supporting the OpenAI Responses and Chat Completions APIs, as well as 100+ other LLMs." The core primitives are agents (models configured with instructions, tools, guardrails and handoffs), agents as tools and handoffs for delegation, guardrails for input and output validation, sessions for conversation history and built-in tracing. A runner executes the agent loop.

AgentKit was the umbrella name OpenAI used at launch for a set of tools around agents. The announcement listed Agent Builder (a visual canvas for multi-agent workflows), ChatKit (an embeddable chat interface), a Connector Registry for managing data and tool connections, Guardrails (an open source safety layer for masking PII and detecting jailbreaks) and new Evals capabilities. The Agents SDK predates AgentKit and is the code-first part of the same stack.

OpenAI's current agents guide compares four ways to build: the Agents API, which runs "a managed Codex harness" and manages session state; the Agents SDK, which "gives your application control over deployment, storage, approvals, and runtime integration"; the Responses API, for direct model calls with your own state and execution; and ChatKit for embedded chat. The same page groups Agent Builder and the Assistants API under "Legacy APIs".

ComponentWhat it isStatus as of 15 November 2026
Agents SDK (Python and TypeScript)Open source library: agents, tools, handoffs, guardrails, sessions, tracingActive; Python 0.24.0, TypeScript 0.22.0
Agent BuilderVisual canvas for designing and versioning workflowsDeprecated; shutdown 30 November 2026
ChatKitToolkit for embedding chat-based agent interfacesAvailable ("ChatKit remains available")
Evals (dashboard, API and graders)Evaluation workflowsRead-only; shutdown 30 November 2026
GuardrailsOpen source safety layer for Python and JavaScriptNot stated by OpenAI
Connector RegistryAdmin hub for data and tool connections across ChatGPT and the APIBeta rollout at launch; later status not stated by OpenAI

Key facts at a glance

As of 15 November 2026, from OpenAI's documentation and GitHub repositories.

FieldAgents SDK for PythonAgents SDK for TypeScript
OwnerOpenAIOpenAI
Packageopenai-agents@openai/agents (with @openai/agents-core and @openai/agents/realtime)
LicenceMITMIT
First tagv0.0.1, 11 March 2025Not stated by OpenAI
Latest release0.24.0 (tagged 15 October 2026); 0.24.x alphas tagged in November 20260.22.0 (package manifest on the main branch)
Runtime requirementPython 3.10 or newerNode.js 22 or later, Deno, Bun; Cloudflare Workers with nodejs_compat is experimental
Installpip install openai-agents (extras: voice, redis, litellm, any-llm)npm install @openai/agents zod
Default modelgpt-5.6-luna since 0.20.0Not stated by OpenAI
VersioningModified semantic versioning in the form 0.Y.Z
PriceFree and open source; model and hosted tool usage billed by the provider

Plans and pricing

The SDK has no licence fee or paid tier. Costs come from what the agent calls. At launch OpenAI said of AgentKit: "All of these tools are included with standard API model pricing." Model tokens are billed at the chosen model's rates on the OpenAI pricing page, and the page states that "Tokens used for built-in tools are billed at the chosen model's per-token rates." The built-in tools that Agents SDK agents can call through the Responses API carry their own charges. Prices below are copied from OpenAI's pricing page as of 15 November 2026, in US dollars.

Hosted toolPrice (OpenAI pricing page)
Web search (including images)$10.00 per 1,000 calls, plus search content tokens at model rates
Web search preview, non-reasoning models$25.00 per 1,000 calls; search content tokens free
Hosted Shell and Code Interpreter containers1 GB $0.03, 4 GB $0.12, 16 GB $0.48, 64 GB $1.92 per 20-minute session per container
File search storage$0.10 per GB per day, first 1 GB free
File search tool call$2.50 per 1,000 calls (Responses API only)
ChatKit file and image upload storage$0.10 per GB-day, first 1 GB free

Other running costs: sandbox agents can run on Docker, a local Unix sandbox or hosted sandbox providers, each billed by that provider. Third-party tracing back ends are billed separately. OpenAI's pricing page does not list a separate price for the Agents API as of 15 November 2026.

Availability: the Agents SDK is open to anyone with an OpenAI API key or a compatible provider. At launch the Connector Registry required the Global Admin Console and rolled out to some API, ChatGPT Enterprise and Edu customers.

Models available

The SDK "uses the Responses API by default for OpenAI models". When an agent sets no model, it uses gpt-5.6-luna with reasoning.effort="none" and verbosity="low", a default the docs describe as intended for "cost-sensitive, high-volume agent workflows". Set OPENAI_DEFAULT_MODEL to change the default for all agents, or pass RunConfig(model=...) for one run. The default has moved several times: gpt-4.1 until 0.16.0, then gpt-5.4-mini, then gpt-5.6-luna from 0.20.0. Realtime agents default to gpt-realtime-2.1 from 0.18.0.

RouteHow it is configuredNotes from the docs
OpenAI Responses APIOpenAIResponsesModel (default)Recommended path for OpenAI models; optional WebSocket transport since 0.10.0
OpenAI Chat Completions APIOpenAIChatCompletionsModel or set_default_openai_api("chat_completions")Use for providers that do not support the Responses API
OpenAI-compatible endpointsAsyncOpenAI client with your base_url and keyGlobal default through set_default_openai_client, or per run through ModelProvider
LiteLLM adapteropenai-agents[litellm], litellm/... model names"best-effort, beta adapter integrations"
Any-LLM adapteropenai-agents[any-llm], any-llm/... model names"best-effort, beta adapter integrations"

Features and modes

  • Agents, handoffs and agents as tools: an agent can hand the conversation to another agent or call it as a tool. From 0.6.0 handoff history is packaged into one assistant message; from 0.7.0 nested handoff history is opt-in through RunConfig(nest_handoff_history=True).
  • Tools: Python functions through @function_tool (with @tool as a shorter alias from 0.19.0), hosted tools such as web search, file search and code interpreter, and Programmatic Tool Calling (0.19.0), where supported models write JavaScript to coordinate eligible tools.
  • MCP: HostedMCPTool lets the Responses API call a remote server; MCPServerStreamableHttp, MCPServerSse and MCPServerStdio connect from your process; MCPServerManager handles several servers. Tools can be filtered with create_static_tool_filter and gated with require_approval.
  • Guardrails and human in the loop: input and output guardrails validate what goes in and out, and approvals can pause a run until a person decides. Since 0.15.0 model refusals raise ModelRefusalError.
  • Sessions and state: automatic conversation history, a Redis option, and resumable RunState objects; 0.20.0 added add_input() to stage input before resuming.
  • Sandbox agents (beta, 0.14.0): SandboxAgent, Manifest and SandboxRunConfig give agents persistent isolated workspaces with files, Git repos, mounts, snapshots and resume. Back ends include UnixLocalSandboxClient, DockerSandboxClient and hosted integrations for Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop and Vercel.
  • Realtime and voice agents: realtime agents for low-latency speech, with SIP support added in 0.5.0, and voice pipelines that chain speech-to-text, an agent workflow and text-to-speech.
  • Tracing: built-in spans for model calls, tools, handoffs and guardrails, viewable in the OpenAI Traces dashboard or exported to other processors.

Supported IDEs and platforms

The SDK is a library, so it runs wherever its runtime runs: Python 3.10 or newer for openai-agents, and Node.js 22 or later, Deno or Bun for @openai/agents, with experimental support for Cloudflare Workers when nodejs_compat is enabled. The TypeScript documentation also lists voice agent extensions for Cloudflare and Twilio and WebRTC and WebSocket transports for realtime sessions. There is no dedicated IDE: the README shows installation with pip or uv, and the repository ships examples under examples/, including sandbox tutorials under examples/sandbox/.

Where the agent runsWhat you manageTypical use from the docs
Your application processHosting, storage, approvals and secretsAgents SDK with the runner in your own service
Local or Docker sandboxThe host or containerDevelopment and containerised sandbox agents
Hosted sandbox providerProvider account and credentialsLong-horizon sandbox agents through optional extras
OpenAI-managed harnessConfiguration through the APIAgents API, which runs a managed Codex harness

Enterprise and admin controls

The SDK itself has no admin console: controls come from the OpenAI platform account behind the API key. Organisation and project scoping are set on the AsyncOpenAI client; from 0.22.0 OpenAIProvider raises UserError if you pass organization or project alongside an explicit client, so configure them in one place. At launch the Connector Registry was described as an admin hub for data and tool connections across ChatGPT and the API, requiring the Global Admin Console. For project roles, service accounts, key scoping and audit logs, see the Securing OpenAI Agents SDK and Responses API Guide.

Data handling and privacy

The tracing documentation states that "Tracing is enabled by default." The table summarises what it records and how to control it.

ItemDefaultControl
Tracing to OpenAIOnOPENAI_AGENTS_DISABLE_TRACING=1, set_tracing_disabled(True), or RunConfig.tracing_disabled per run
Model and tool inputs and outputs in spansIncluded (trace_include_sensitive_data defaults to true)RunConfig.trace_include_sensitive_data or OPENAI_AGENTS_TRACE_INCLUDE_SENSITIVE_DATA=false
Audio in voice spansBase64 PCM audio includedVoicePipelineConfig.trace_include_sensitive_audio_data
Other trace destinationsOpenAI exporter onlyadd_trace_processor() adds a processor; set_trace_processors() replaces the defaults
Zero Data Retention organisationsTracing unavailableNot applicable

The docs add that trace processors are independent, so if a redaction processor fails, others still receive data; put redaction and delivery in one exporter when export depends on redaction. They list third-party integrations such as Weights & Biases, Arize Phoenix, MLflow, Langfuse and LangSmith, and state that the listing is not an endorsement or security certification. Release 0.19.0 hardened error and diagnostic logging to avoid exposing raw sensitive payloads.

Release history

Dates are GitHub tag dates for the Python repository, with changes taken from the SDK's breaking change log, as of 15 November 2026.

DateReleaseWhat changed
11 Mar 2025v0.0.1First tagged release
27 Jun 20250.1.0MCPServer.list_tools() gains run_context and agent
11 Sep 20250.3.0Realtime support moves to the GA gpt-realtime interface
6 Oct 2025AgentKit announcedAgent Builder (beta), ChatKit, Connector Registry, Guardrails, Evals
18 Oct 20250.4.0Requires openai v2
18 Nov 20250.6.0Handoff history packaged into one message
22 Jan 20260.7.0Nested handoff history opt-in; default reasoning effort "none" for gpt-5.1 and gpt-5.2
13 Feb 20260.9.0Python 3.9 support dropped
23 Feb 20260.10.0WebSocket transport for the Responses API (opt-in)
15 Apr 20260.14.0Sandbox Agents (beta)
7 May 20260.16.0Default model gpt-5.4-mini; max_turns=None allowed
3 Jun 2026Deprecation noticeAgent Builder deprecated and Evals read-only, both with shutdown on 30 November 2026
27 Jul 20260.19.0Programmatic Tool Calling; @tool alias
11 Aug 20260.20.0Default model gpt-5.6-luna; MCP Python SDK v2 support
15 Aug 20260.21.0Requires openai v3 and HTTPX2
19 Aug 20260.22.0Output guardrail withholding; stricter provider argument checks
15 Oct 20260.24.0Latest minor tag at the time of checking
30 Nov 2026ShutdownAgent Builder and Evals scheduled to shut down

Versioning policy: the release page says the leading 0 "indicates the SDK is still evolving rapidly". Minor versions (Y) are increased "for breaking changes to any public interfaces that are not marked as beta", and patch versions (Z) cover bug fixes, new features, private interface changes and updates to beta features. Some minors, such as 0.14.0, 0.18.0 and 0.19.0, are labelled as introducing no breaking change.

Limits and known constraints

  • Agent Builder workflows need a plan before 30 November 2026: the deprecation names the Agents SDK or ChatGPT Workspace Agents as replacements. Evals is read-only with the same shutdown date, and OpenAI's recommended replacement is to "Migrate to Promptfoo".
  • Default model changes alter cost and behaviour: agents without an explicit model changed model at 0.16.0 and again at 0.20.0. Set the model explicitly in production.
  • Dependency jumps: 0.21.0 requires openai v3 and moves HTTP integrations to HTTPX2; custom httpx clients and transports must be migrated.
  • Beta areas: sandbox agents and the LiteLLM and Any-LLM adapters are labelled beta, and the SSE MCP transport is deprecated by the MCP project.
  • Tracing and ZDR: organisations under a Zero Data Retention policy cannot use OpenAI tracing and need another processor for observability.
  • Language parity: some features appear in Python first; sandbox hosted provider integrations are listed as optional extras in the Python package.

Security guidance

This page covers what the SDK and AgentKit are, how they are versioned and what they cost. Hardening is covered in the Securing OpenAI Agents SDK and Responses API Guide, which goes through API keys, service accounts, hosted MCP approvals and audit logs. From the SDK's own documentation, the most useful built-in controls are require_approval on MCP servers and approvals for function tools, static or dynamic tool filters, input and output guardrails, and the sandbox manifest rules: release 0.17.0 keeps local sources inside the materialisation directory unless covered by Manifest.extra_path_grants, and the docs say not to populate grants from model output. For credentials inside sandboxes, 0.20.0 rejects unsafe credential placement in mounts. Apply OpenAI security updates promptly. The wider discipline is covered in the Securing Agent Frameworks Guide and the Human-in-the-Loop for AI Agents Guide.

Practitioner checklist

  • List every Agent Builder workflow and Evals dependency and move them before the 30 November 2026 shutdown.
  • Pin openai-agents or @openai/agents to a minor version and read the breaking change log before each minor upgrade.
  • Set the model explicitly on every production agent instead of relying on the SDK default.
  • Decide on tracing per environment: disable it, set trace_include_sensitive_data to false, or route traces to your own processor.
  • Confirm the openai v3 and HTTPX2 migration before taking 0.21.0 or later if you customise HTTP clients.
  • Gate destructive MCP and function tools with require_approval and filter exposed tools to what the agent needs.
  • Budget hosted tool charges (web search, file search, containers) separately from model tokens.
  • Treat sandbox path grants and mount credentials as trusted configuration, never as model output.

Standards and references

Related NHI Mgmt Group resources: Securing OpenAI Agents SDK and Responses API Guide · Securing Agent Frameworks Guide · AI Agent Sandboxing and Code Execution Guide · Human-in-the-Loop for AI Agents Guide

PDF for subscribers

Take this guide with you

Every section, table and checklist in one PDF, with links to the sources and related guides. Subscribe to our weekly NHI & AI Identity Journal and we’ll email you a download link.

Free weekly newsletter

Subscribe to the NHI & AI Identity Journal

The latest on NHI and Agentic AI security – articles, research, breaches, news and events every week.

Bonus 33% off our NHI Course when you subscribe.

NHIMG Editorial Note
Researched and written by the NHI Mgmt Group editorial team using AI tools, under the editorial direction of Lalit Choda. Facts are drawn from the sources listed on this page and checked by our automated quality process. Last updated 15 November 2026. Spotted an error? Use our contact page.
NHI Mgmt Group, the independent authority on Non-Human Identity, IAM and Agentic AI security. nhimg.org

PDF Downloads for Newsletter Subscribers

Get the guide

Enter your details and we’ll email you a download link.

The full guide is free to read online. The PDF is a thank you for subscribing.

Fill in all fields marked * and tick the PDF box to continue.