OpenAI Agents SDK−17 Mixed

GILD Review

GILD's call and the buying facts behind it.

From GILD research

Should we adopt OpenAI Agents SDK?

Should your team build its agents on the OpenAI Agents SDK, a graph framework, or its own loop?

30-Second TakePick the OpenAI Agents SDK if your agents run on OpenAI models and need handoffs and approvals, with no graph framework.
Best fitTeams building agents on OpenAI models in their own Python or TypeScript service, who want handoffs, guardrails, and approvals without a graph framework.GILD Review
Consider another workflowFor explicit, checkpointed control flow across model vendors, evaluate LangGraph. For one TypeScript code path across many models with UI hooks, try the Vercel AI SDK, and for Claude agents that work in files and a shell, the Claude Agent SDK. For a harness OpenAI hosts, OpenAI's Agents API is in public beta.GILD Review

Drawn from 75 sources and 10 operator insights

GILD analysis

OpenAI Agents SDK Review

Harvey builds its agentic search evals on the OpenAI Agents SDK's traces, and small companies from Lark to Octagon AI ship products on it. The small MIT-licensed Python or TypeScript library gives a team agents, tools, handoffs, guardrails, sessions, and approvals in its own service. With no model set it runs gpt-5.6-luna, and its traces, prompts and tool data included, go to OpenAI by default.

Agent Frameworks
Agentic Development

The decision

Best fit: Teams building agents on OpenAI models in their own Python or TypeScript service, who want handoffs, guardrails, and approvals without a graph framework.
Consider another workflow: For explicit, checkpointed control flow across model vendors, evaluate LangGraph. For one TypeScript code path across many models with UI hooks, try the Vercel AI SDK, and for Claude agents that work in files and a shell, the Claude Agent SDK. For a harness OpenAI hosts, OpenAI's Agents API is in public beta.
Adoption condition: A pinned SDK version and model, a decision on where traces go, shared session storage, and guardrails and approvals tested before the agent reaches users.

Basis: OpenAI's Agents SDK documentation, API pricing, model, and data-control pages checked Oct 28, 2026, plus 59 public accounts from builders and teams, Harvey's engineering blog among them. We ran no performance benchmark. Buying facts rechecked Oct 31, 2026.

Buying & deployment factsWhat a team building on it should verify before production traffic grows.

How you buy it

Free, MIT license

Model calls bill to an OpenAI API account, with gpt-5.6-luna as the default model.

Source 1, OpenAI, openai-agents-python on GitHub: openai-agents-python on GitHub Source 2, OpenAI, Agents SDK models: Agents SDK models

Details

The SDK is free and open source under the MIT license, installed with pip install openai-agents. There is no SDK plan or fee: model calls bill to the OpenAI API account whose API key the SDK uses. The default model is gpt-5.6-luna when an agent sets none, and OPENAI_DEFAULT_MODEL or RunConfig changes it.

Cost drivers

US$0.20 in, US$1.20 out

Per million tokens on gpt-5.6-luna, pay-as-you-go. Build tier usage caps at US$500 a month.

Source 3, OpenAI, gpt-5.6-luna model page: gpt-5.6-luna model page Source 4, OpenAI, OpenAI API rate limits and usage tiers: OpenAI API rate limits and usage tiers

Details

Tokens on every turn, tool call, and handoff, billed pay-as-you-go per million tokens at OpenAI API prices. The default gpt-5.6-luna lists US$0.20 per million input tokens and US$1.20 per million output tokens. Usage tiers rise with total credit purchases, and each tier caps monthly usage, at US$500 a month on Build.

Where it runs

Your own process

Python 3.10 or newer. Other models run through beta LiteLLM and Any-LLM adapters.

Source 1, OpenAI, openai-agents-python on GitHub: openai-agents-python on GitHub Source 2, OpenAI, Agents SDK models: Agents SDK models

Details

In your own process, on Python 3.10 or newer. For OpenAI models OpenAI recommends its Responses model, which calls the Responses API. Other models run through the LiteLLM and Any-LLM adapters, which OpenAI supports on a best-effort, beta basis.

What to verify

5,000 RPM on Build

gpt-5.6-luna is not supported on the Free tier. Test each model you plan to run.

Source 3, OpenAI, gpt-5.6-luna model page: gpt-5.6-luna model page Source 2, OpenAI, Agents SDK models: Agents SDK models Source 5, trust.openai.com, OpenAI Trust Portal: OpenAI Trust Portal Source 6, OpenAI, OpenAI API data controls: OpenAI API data controls

Details

gpt-5.6-luna is not supported on the Free usage tier and starts at 5,000 RPM on the Build tier. OpenAI says its products support SAML. OpenAI says its API platform products are ISO 27001, 27017, 27018, and 27701 certified. OpenAI's Admin and Audit Logs API lets API organization owners and admins modify and document user permissions to support audit logging. By default, OpenAI keeps abuse monitoring logs for up to 30 days, with longer retention in cases OpenAI lists.