Event pulse · Agents API public beta

Not affiliated with OpenAI

Agents API sandbox choices: hosted vs Cloudflare, DO, Vercel

Q: Does the Agents API add a harness surcharge?

No. OpenAI’s September 10, 2026 announcement says there are no additional fees for using the Agents API — you pay for the tokens and tools your agents use. Docs add that OpenAI-hosted sandboxes are billed at standard container rates. Partner or self-hosted compute is billed by that provider or by you.

Q: Can I run the sandbox on Cloudflare, DigitalOcean, or Vercel?

Yes. The Agents API lets you pick OpenAI-hosted compute or a self_hosted environment. Official partner guides cover Cloudflare, DigitalOcean, Vercel, E2B, Modal, Daytona, Blaxel, Oracle Cloud Infrastructure, and Runloop. Self-hosted can also be your own laptop, Docker, or other infrastructure.

Q: Does a self-hosted sandbox give US-only residency a pass, or enable Zero Data Retention?

No. OpenAI’s Agents API overview states the API currently supports data residency only in the United States and does not support Zero Data Retention (ZDR). Choosing a self-hosted sandbox does not make the Agents API ZDR-eligible.

Q: Is Agent Builder the same as the Agents API?

No. Agent Builder is a separate product scheduled to shut down on November 30, 2026. OpenAI’s deprecations page points migrations to the Agents SDK or ChatGPT Workspace Agents. The Agents API is a public-beta managed Codex harness (sessions, compaction, tools, sandboxes), billed at model/tool/container rates — not a replacement calculator for Agent Builder graphs.

Q: If the agent uses GPT-6 Astra, what token rates apply?

Model usage is billed at the selected model’s API rates. For GPT-6 Astra, OpenAI’s Standard card last checked October 13, 2026 is $10 input, $1 cached input, $12.50 cache writes, and $50 output per 1M tokens, with Batch and Flex at half, Fast at 2×, Ultrafast at the published $60/$6/$75/$300 short-context card, and higher long-context rates above 272K input tokens. Confirm on OpenAI’s model docs and pricing page.

This is a cost-shape explainer, not a sandbox-hour calculator. OpenAI runs the Codex harness. You pick where code executes — OpenAI-hosted, a partner sandbox, or your own machines — and the bill splits along that cut.

If the agent’s model is GPT-6 Astra, token lines still follow Astra API rates (Standard / Batch / Flex / Fast, plus the 272K cliff). Run that part here; confirm sandbox SKUs on OpenAI and the partner.

Open the Astra pricing calculator

What shipped

On September 10, 2026 OpenAI put the Agents API in public beta: the same Codex harness that coordinates model calls, tools, and context, exposed as POST /v1/agents/sessions (beta header agents=v1).

Durable sessions — session state is retained so you continue work across turns without rebuilding context. Delete sessions (and published artifacts) when you are done.
Context compaction — the harness compacts earlier context as a session nears its window so workflows can span multiple windows without your own compaction loop.
Tools / MCP — MCP servers, custom functions, built-in tools (e.g. web search), tool search, and programmatic tool calling (parallel / chained calls in code).
Subagents — multi_agent can fan work to concurrent subagents with their own context; the parent coordinates.
Environment — none, openai_hosted, or self_hosted (your laptop, Docker, or a partner sandbox).

Sources: Introducing the Agents API, Agents API overview.

What you pay for

OpenAI’s launch post is explicit: there are no additional fees for using the Agents API — you pay for the tokens and tools the agent uses. Docs add the third line when OpenAI owns the box:

Model tokens — billed at the selected model’s API rates. Pick gpt-6-astra and you are on Astra token pricing (see Astra API cost and the calculator).
Tools — OpenAI tools at their published tool rates (web search, file search, and so on). Tokens those tools feed the model still count as model usage.
OpenAI-hosted sandbox — standard container rates, billed separately from the model. Official pricing, last checked October 13, 2026, still lists Hosted Shell / Code Interpreter containers in 1 / 4 / 16 / 64 GB memory tiers at $0.03 / $0.12 / $0.48 / $1.92 per 20-minute session per container. Eligible container sessions are billed by the minute with a 5-minute minimum. Re-read the live card before you budget; this page is not a container-minute calculator.
Partner or self-hosted sandbox — no OpenAI container line for that compute. You pay the partner (or your own infra) on their SKU. The Agents API still bills tokens and OpenAI tools. Idle compute you forget to tear down is a partner/ops problem, not an OpenAI harness surcharge.

Sources: launch post (no extra API fee), overview pricing blurb, hosted sandbox “standard container rates”, API pricing → Containers, community thread. Community staff repeated: hosted sandboxes use standard container rates; model usage is separate. Confirm numbers on OpenAI’s page — rates move.

Sandbox choice matrix (qualitative)

OpenAI hosts the harness either way. The cut is who provisions the Linux workspace and whose meter runs while it is up. Partner rows are first-class self_hosted integrations from OpenAI’s environment docs — not a ranking, and not a $/hour table (those SKUs live on each vendor’s bill).

Option	Who runs compute	Cost shape	Fit
OpenAI-hosted	OpenAI provisions Linux (Python, Node, CLI). Working dir `/workspace`.	OpenAI container rates + model + tools. Files persist across turns until the sandbox expires; idle ~1 hour without keep-alives can delete it. Outputs under `/workspace/outputs` publish as artifacts.	Fastest path. You do not bring an image, VPC, or partner account. Network can be enabled / disabled / host-allowlisted. Templates save config, not a running VM.
Cloudflare	You (or a Worker) run a Cloudflare Container per session; executor connects outbound.	Cloudflare Worker + Container usage on Cloudflare’s bill. OpenAI still bills tokens/tools. Reference Worker shuts the container down when idle.	Already on Workers; want the workspace inside your Cloudflare account; signed webhooks for session orchestration.
DigitalOcean	DO M.A.R.S. starts a Firecracker microVM from the `codex-agentapi` image; executor connects outbound.	DigitalOcean sandbox/runtime bill + OpenAI tokens/tools. Needs a sandbox-enabled DO account with access to that image.	Want a microVM boundary and DO-side session mapping (`openai_session_id`). Webhook- or application-managed provisioning.
Vercel	Vercel Sandbox as the isolated executor; app/Functions + queues handle signed webhooks.	Vercel Function / Queue / Sandbox SKUs + OpenAI tokens/tools. Sandbox has no inbound port — executor dials out.	App already on Vercel; want a persistent isolated sandbox per session with webhook-driven lifecycle.
Other partners (E2B, Modal, Daytona, Blaxel, Oracle/OCI, Runloop)	Provider sandbox or devbox; same outbound executor pattern.	That vendor’s compute/storage SKU + OpenAI tokens/tools. Pause vs replace matters (e.g. E2B: pause keeps files; replace does not).	Need GPU/memory options, VPC/deploy-in-your-cloud, snapshots, or mounts those vendors document. OpenAI lists first-class guides for each.
Self-hosted / DIY	Laptop, Docker, or any box running `codex exec-server`.	Your electricity/cloud VM bill. OpenAI tokens/tools only on the OpenAI invoice. You own cleanup.	Private network, custom image, air-gapped files. Restricted executor key in the box; application key stays out. Does not change residency/ZDR (see below).

Environment types: OpenAI-hosted, self-hosted + partner list. Cloudflare: Sep 10 changelog, tutorial. DigitalOcean: DO provider guide. Vercel: Vercel KB. Launch partner list also names Blaxel, Daytona, E2B, Modal, Oracle, Runloop.

Beta caveat: US residency, no ZDR

OpenAI’s Agents API overview states the API currently supports data residency only in the United States and does not support Zero Data Retention (ZDR). Choosing a self-hosted or partner sandbox does not make the Agents API ZDR-eligible. Session state is retained so the agent can resume; you can delete sessions and published artifacts when you no longer need them (physical cleanup may continue asynchronously).

If the workload is contractually non-US-at-rest or ZDR-only, the current public-beta Agents API is the wrong product — that is a policy gate, not a sandbox-vendor tweak. Platform-wide residency and retention rules live in OpenAI’s data-controls docs; do not assume a Cloudflare/DO/Vercel box changes the API’s residency story.

Sources: Agents API overview (US-only, no ZDR, self-hosted ≠ ZDR), manage sessions / delete, Data controls in the OpenAI platform.

Agent Builder / Evals: Nov 30, 2026

Separate from this beta, OpenAI is retiring Agent Builder and the Evals dashboard/API. Deprecation was announced June 3, 2026. Agent Builder is scheduled to shut down November 30, 2026 (ChatKit stays). Evals become read-only October 31, 2026 and are scheduled to shut down November 30, 2026.

Official migration: export Agent Builder workflows to the Agents SDK (TypeScript or Python) or rebuild as ChatGPT Workspace Agents. Evals: OpenAI points at Promptfoo / repo-owned tests. That is not “paste the graph into the Agents API” — the API is a managed Codex session runtime; the SDK is what the deprecations page names for code.

Sources: Deprecations, Migrate from Agent Builder, AgentKit post (June 3, 2026 update).

FAQ

Does the Agents API add a harness surcharge?

No. The September 10, 2026 announcement says there are no additional fees for using the Agents API — you pay for the tokens and tools your agents use. Docs add that OpenAI-hosted sandboxes use standard container rates. Partner/self-hosted compute is on that provider’s (or your) invoice.

Can I run the sandbox on Cloudflare, DigitalOcean, or Vercel?

Yes. Set environment.type to self_hosted and connect an executor. OpenAI’s partner list includes Cloudflare, DigitalOcean, Vercel, E2B, Modal, Daytona, Blaxel, Oracle Cloud Infrastructure, and Runloop. Same pattern: outbound codex exec-server, restricted environment key inside the box, application key outside.

Does a partner sandbox fix US-only residency or enable ZDR?

No. The Agents API overview says residency is United States only and ZDR is not supported. A self-hosted sandbox does not make the API ZDR-eligible.

Is Agent Builder the same as the Agents API?

No. Agent Builder shuts down November 30, 2026; migrate via Agents SDK or ChatGPT Workspace Agents. The Agents API is the public-beta managed Codex harness (sessions, compaction, tools, sandboxes).

If the agent uses GPT-6 Astra, what token rates apply?

Model usage is billed at that model’s API rates. Astra Standard, last checked October 13, 2026: $10 input, $1 cached input, $12.50 cache writes, $50 output per 1M, with Batch/Flex at half, Fast at 2×, and Ultrafast at the published $60 / $6 / $75 / $300 short-context card. Long prompts use the 272K cliff. Use the calculator or the API cost page; verify on OpenAI’s GPT-6 Astra docs and API pricing.

Official docs

Introducing the Agents API (Sep 10, 2026)
Agents API overview
OpenAI-hosted sandboxes
Self-hosted sandboxes / partner list
API pricing (tokens, tools, containers)
Community: Agents API and hosted sandboxes
Deprecations (Agent Builder / Evals)

Related on this site

Astra pricing calculator (homepage — GPT-6 token estimate)
Astra API cost — token lines & tiers
Astra 272K pricing — long-context cliff