OpenAI's Agents API went into public beta on 10 September 2026, and its pricing fits in one sentence from the overview: "Model usage is billed at the selected model's API rates. OpenAI tools use their standard rates, and OpenAI-hosted sandboxes use standard container rates."
So there is no Agents API fee. A session costs three things, and this page puts official numbers on each. All figures are from OpenAI's pricing page, checked on 28 September 2026.
Disclosure: we build Gobare, a hosted agent runtime that runs on your own model key. The last section says how the bill splits there and on the other alternatives.
1. The container
An OpenAI-hosted sandbox is billed per container, per 20-minute session:
| Memory | Price per 20-minute session |
|---|---|
| 1 GB | $0.03 |
| 4 GB | $0.12 |
| 16 GB | $0.48 |
| 64 GB | $1.92 |
Hosted Shell and Code Interpreter share these rates. A self-hosted sandbox or a sandbox partner moves this line to your own bill or the partner's.
2. The model
Tokens are billed at the chosen model's rates. The two ends of the GPT-6 range, per million tokens, short context:
| Model | Input | Cached input | Output |
|---|---|---|---|
| gpt-6-astra | $10.00 | $1.00 | $50.00 |
| gpt-6-luna | $0.10 | $0.01 | $0.50 |
gpt-6-astra is the model every example in the Agents API docs uses. The
difference between the two is a factor of a hundred.
3. The tools
Only if the agent uses them:
- Web search: $10.00 per 1,000 calls, plus search content tokens at model rates.
- File search: $2.50 per 1,000 calls, and storage at $0.10 per GB per day after the first free GB.
A worked example
One task: the agent reads a small repository, makes a change, runs the tests. Say it takes one 20-minute container window at 4 GB, consumes 60,000 input tokens and produces 10,000 output tokens, with no tool calls. This is arithmetic on the published rates, not a measurement:
| gpt-6-astra | gpt-6-luna | |
|---|---|---|
| Input tokens (60k) | $0.60 | $0.006 |
| Output tokens (10k) | $0.50 | $0.005 |
| Container (4 GB, one window) | $0.12 | $0.12 |
| Total | $1.22 | $0.131 |
Two things follow. On the expensive model, tokens are most of the bill and the container is noise. On the cheap model, the container is over 90% of the bill — so the sandbox, not the model, is what you are really pricing. That is also the point at which model choice stops being a line item and starts being a reason to look elsewhere.
How the bill splits on the alternatives
Every hosted alternative splits the same three costs differently. The ones with published numbers:
- Bring-your-own-key runtimes send token costs to your model provider at that provider's rates, and charge only for the runtime. On Gobare the key is yours and there is no margin on tokens; the compute rate is not published yet and the service is free while in alpha.
- DigitalOcean Managed Agents bills per second of active CPU. Its own example: two vCPUs averaging 25% utilisation with a 4 GB memory peak for an hour costs $0.060 in CPU and memory.
- Perplexity's sandbox tool is $0.03 per session, where a session "covers up to 20 minutes of active use for billing purposes".
- Self-hosted options (Open Managed Agents, AstraBox) move the runtime and the sandbox onto infrastructure you already pay for.
Which of these is cheaper depends on the shape of your work: short tasks on a cheap model favour per-second billing, long idle waits favour anything that does not bill while paused.
The full comparison — what each alternative keeps, which ones accept the
official openai package unchanged, and where each one loses — is in
OpenAI Agents API alternatives in 2026.
Frequently asked questions
Does the OpenAI Agents API cost extra on top of tokens?
No separate fee. You pay model tokens, the container if you use an OpenAI-hosted sandbox, and any OpenAI tools the agent calls.
How much does an OpenAI-hosted sandbox cost?
From $0.03 (1 GB) to $1.92 (64 GB) per 20-minute session per container.
What is the cheapest way to run the OpenAI Agents API?
Use a small model such as gpt-6-luna and the smallest container that fits the task. At that point the container is most of the bill, which is worth knowing before you compare runtimes.
Can I use my own model key to lower the cost?
Not on the hosted Agents API, which runs OpenAI models. Runtimes that take your own key, including Gobare, bill tokens through your provider instead — see the alternatives.
Sources, checked 28 September 2026: OpenAI pricing, Agents API overview; DigitalOcean Managed Agents; Perplexity pricing and sandbox tool.