OpenAI AgentKit → LangGraph
What it saves, what actually moves, what you rebuild — and the thing that catches people.
What it costs, and what it saves
Roughly neutral on model spend — you still pay per token. What you avoid is a product being withdrawn: OpenAI announced on 3 June 2026 that Agent Builder is winding down and leaves the platform on 30 November 2026.
Moves cleanly
Prompts, tool definitions and model calls. Those are yours.
You rebuild
The orchestration. A graph built in Agent Builder lives in OpenAI's product, not in a file you own, so it is rebuilt rather than exported.
What OpenAI AgentKit costs you today
There was no separate subscription — you paid ordinary API rates for what the agent consumed, roughly $1.75 per million input tokens and $14 per million output tokens on the flagship model, plus tool charges such as $0.03 per Code Interpreter session and $0.10 per gigabyte per day for file search. Enterprise agreements sat between roughly $2,000 and $20,000 a month. Figures checked 2026-07-30. The pricing was never the problem.
What actually holds you in
The prompts, the tool definitions and the model calls are portable — those are yours and they move. What does not move is the orchestration: a graph built in Agent Builder is a description held in OpenAI's product, not a file in your repository, and rebuilding it elsewhere means rebuilding the logic rather than exporting it. Evals histories are similar. This is the specific shape of visual-builder lock-in: the easier a platform makes it to avoid writing the wiring down, the less of the wiring you own when it leaves.
What you are moving to
LangGraph models an agent as a graph of nodes with explicit state, checkpointing and human-in-the-loop interrupts, which is very nearly the mental model Agent Builder gave you — with the difference that the graph is code in your repository. It is the most battle-tested option for production agents that must survive restarts, resume mid-run and be inspected when they go wrong. MIT-licensed, roughly 38.5k GitHub stars and pushed to daily as of 2026-07-30. LangGraph Platform is a paid hosted deployment tier, but the framework itself is complete without it — the distinction is worth checking before you build on anything you assume is free.
Free and open source. LangGraph Platform is an optional paid hosted tier.
LangGraph strengths
- Explicit state graph maps closely onto what Agent Builder did
- Checkpointing and resumable runs — genuinely production-grade
- Human-in-the-loop interrupts are a first-class feature, not a workaround
- MIT and in your repository: nobody can announce a sunset date for it
What you give up
- Steepest learning curve here — you are writing the graph, not drawing it
- Carries LangChain ecosystem conventions whether or not you want them
- Verbose for simple agents that only need a loop and two tools
- The hosted platform is where the operational conveniences live
The migration, step by step
- 1Screenshot or document every Agent Builder workflow now, while the product still exists
- 2Model each as a LangGraph StateGraph — nodes for steps, edges for transitions
- 3Port tool definitions, which map closely
- 4Add checkpointing so runs survive restarts; this is something Agent Builder did invisibly
- 5Test against the same inputs and compare outputs before the November deadline
The gotcha
The deadline is real and there is no extension. Documenting the workflows before the product is switched off is the step people skip, and afterwards there is nothing left to read.
When to stay on OpenAI AgentKit
Nothing to stay for — it is being switched off. Migrate before November 2026.
More Layer 4 migrations
Figures verified 2026-07-30 against vendor pricing pages. Prices change and migrations differ by estate — treat the cost delta as a starting model, not a quote. Rankings and recommendations here are merit-only; affiliate income never changes a verdict. See our methodology.