Table of Contents
OpenAI Agents API vs Anthropic Managed Agents: Pricing, Limits, and Which to Use (September 2026)
Pick on data boundary and cost-control mechanism, not on sticker price, because on the same workload the two land within about 7% of each other. If any of your agent data is regulated or has to stay outside the US, OpenAI Agents API is off the table today: it is US-only with no Zero Data Retention at launch (per MarkTechPost's coverage of the OpenAI announcement). If you need a hard ceiling on what a single agent run can spend before it pages you at 2 a.m., Anthropic Managed Agents gives you a per-session dollar cap and OpenAI does not.
This is the managed-cloud tier of the decision, where you pay a vendor to run the agent loop for you. The framework tier, Claude Agent SDK vs OpenAI Agents SDK vs build-your-own, is a different question with different economics, covered in the enterprise agent harness decision. Here the loop, the sandbox, and the durable-execution machinery are the vendor's problem. What you are actually buying is a pricing structure and a set of limits. Both opened up in the last few weeks: OpenAI shipped the Agents API in public beta on September 10, 2026 (per the launch coverage), and Anthropic added session budget caps to Managed Agents over the summer. The table decides most of it.
The decision table
| Axis | OpenAI Agents API | Anthropic Managed Agents |
|---|---|---|
| Pricing model | Tokens + tools + container time, no platform fee (per MarkTechPost) | Model tokens + web search at $10 / 1,000 + session runtime at $0.08 / session-hour (per Verdent's pricing breakdown) |
| Data residency | US-only at launch (per MarkTechPost) | No published US-only restriction; treat as a commercial-terms question with your account team |
| Zero Data Retention | Not supported at launch (per MarkTechPost) | Governed by your Anthropic enterprise agreement, not blocked at the product level |
| Sandbox options | OpenAI-managed sandbox, self-hosted codex exec-server, or one of 9 partner sandboxes (per MarkTechPost) | One managed sandbox built into the service; no partner marketplace |
| Session cost control | Org-level spend limits and tier rate limits only; no per-session dollar cap | Per-session dollar cap, checked between model requests (per TokenCost) |
| MCP support | Yes, plus custom functions and built-in tools (per MarkTechPost) | Yes |
Read it as a set of gates. Two rows carry hard constraints that can settle the whole thing on their own: data residency (a regulated or non-US workload rules OpenAI out) and session cost control (a runaway-spend risk you cannot absorb rules OpenAI out). Everything below the table is confirmation of what those two rows already told you.
What OpenAI Agents API handles, and what you still own
The Agents API wraps the harness behind Codex into four concepts: an agent (model, instructions, tools, MCP servers), an environment (an optional sandbox), a session (a durable agent instance), and the events and items flowing through it (per MarkTechPost). OpenAI runs session management, context-window compaction, and recovery as limits approach. You supply the tools and pick where the code runs.
Sandbox choice is the standout. You get three modes: an OpenAI-managed sandbox, a self-hosted option over a WebSocket via codex exec-server, or one of nine partner sandboxes with first-class integrations (Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel) (per MarkTechPost). That partner list is the escape hatch from a pure black box: you can run the execution environment on infrastructure you already pay for and audit.
What you still own is compliance posture and spend governance. The hard limit is the one your security team will hit first: US-only residency and no ZDR at launch means regulated data, or any data that has to stay in-region, cannot go through it yet. Cost control is org-level spend limits and tier rate limits, not a per-run ceiling, so a single misbehaving session is bounded only by your whole org's cap.
The escape path is decent. Because the tool layer speaks MCP and the sandbox can run on a partner like DigitalOcean or Cloudflare, moving off the OpenAI-managed sandbox is a config change, not a rewrite. Moving off the harness itself is harder, since you never ran the loop.
What Anthropic Managed Agents handles, and what you still own
Managed Agents runs the loop in Anthropic's infrastructure with a built-in sandbox where Claude reads files, runs commands, browses the web, and runs code, plus prompt caching and compaction. Pricing is three lines: model tokens at standard Claude rates, web search at $10 per 1,000 searches ($0.01 each), and session runtime at $0.08 per session-hour that accrues only while the session status is running (per Verdent). A 20-minute active run costs about $0.027 in runtime, so the runtime line is usually a rounding error next to tokens.
The feature that separates it is the session budget cap. You set a dollar ceiling per session, and the platform checks it between model requests rather than mid-request (per TokenCost). A request already in flight when the cap is crossed runs to completion, so a session capped at one figure can settle slightly above it, but the overshoot is bounded by a single request, not an open loop. For a platform team, that is the difference between a runaway agent costing you a few extra cents and costing you the monthly budget.
What you still own is observability and portability. The debugger is where this breaks first: when a long session fails halfway through, your visibility is bounded by what the managed harness surfaces, because you do not run the loop. The second limit is model coupling. The harness is wired to Claude, so a business reason to route to a different model means rebuilding the agent, not flipping a config value. The escape path is a rebuild, which is the price you accept for not operating any of the substrate.
What 1,000 sessions actually cost
Both platforms bill web search at $10 per 1,000 calls (OpenAI and Anthropic via Verdent, both fetched 2026-09-17), and both bill model tokens at standard API rates. So the workload lands in the same neighborhood on both. Price the same synthetic job on each: 1,000 research sessions, 200,000 tokens per session, 20 web searches per session, roughly 10 minutes of active runtime each.
Example: 1,000 sessions, 200K tokens each (assume 160K input / 40K output),
20 web searches, ~10 min active runtime. Per-unit rates fetched 2026-09-17.
Anthropic Managed Agents (Claude Sonnet 4.6, $3 input / $15 output per MTok)
input 0.160 MTok x $3.00 = $0.480
output 0.040 MTok x $15.00 = $0.600
web 20 x $0.01 = $0.200
runtime (10/60) h x $0.08 = $0.013
per session ~ $1.29
x 1,000 sessions ~ $1,293
OpenAI Agents API (gpt-5.6-terra, $2.50 input / $15 output per MTok)
input 0.160 MTok x $2.50 = $0.400
output 0.040 MTok x $15.00 = $0.600
web 20 x $0.01 = $0.200
container billed by the minute, 5-min minimum (small)
per session (tokens + web) ~ $1.20
x 1,000 sessions ~ $1,200 + container timeModel rates are the published September 17, 2026 numbers: Sonnet 4.6 at $3 / $15 per million tokens (per Verdent), gpt-5.6-terra at $2.50 / $15 (per OpenAI). OpenAI container sessions are billed by the minute with a five-minute minimum (per OpenAI), a small addition on a ten-minute run. The token split is an assumption; swap in your own and the shape holds.
The finding is not the small dollar gap between the two runs above. It is that price is a wash and the two things that are not a wash are the residency block and the cost-control mechanism. OpenAI can cap your org, not your run. Anthropic can cap the run.
The flip conditions
Each platform owns one case outright.
OpenAI Agents API wins when you want the execution sandbox on infrastructure you already control. No competitor ships nine first-class partner sandboxes. If your platform team already runs Cloudflare or a DigitalOcean fleet and wants the agent's code execution to land there, on machines you audit and pay for directly, OpenAI is the only managed harness that lets you do it without self-hosting the whole loop.
Anthropic Managed Agents wins when a single runaway session is an unacceptable risk. If you are exposing agents to untrusted input, or running thousands of sessions where one bad loop could burn the budget, the per-session dollar cap is the control OpenAI has no answer for today. Org-level spend limits do not protect you from one session spending like a hundred.
Which one to pick, by org profile
US-only startup, no regulated data: OpenAI Agents API. The residency limit does not touch you, pricing is competitive, and the partner-sandbox flexibility is a real edge. Wire spend governance at the org level and set conservative tier limits, since you have no per-session cap to fall back on.
Regulated or global shop: Anthropic Managed Agents now, or wait for OpenAI ZDR. OpenAI's US-only-plus-no-ZDR launch posture blocks regulated and non-US workloads outright (per MarkTechPost). Treat residency and ZDR on the Anthropic side as a commercial-terms conversation with your account team, get it in the DPA, and you can ship this quarter instead of waiting on OpenAI's roadmap.
Cost-sensitive team running high session volume: Anthropic Managed Agents. The per-session budget cap gives you a tighter, per-run ceiling than org-level limits, and the runtime line ($0.08 per session-hour, active time only, per Verdent) is negligible. When your finance team asks what the worst-case cost of one agent run is, you want a number you set, not a story about org caps.
None of these is a default for everyone. Two constraints do the deciding here, same as they did at the table: where your data is allowed to live, and how tight a per-run cost ceiling you need. Answer those two honestly and the profile you fall into is usually obvious.
Wiring it into your stack
Neither platform is where your existing automation lives, so most teams call them from an orchestrator. Both n8n and Make can invoke either service from an AI agent step, which lets you keep triggers, retries, and human-in-the-loop approvals in a tool your team already runs. For n8n specifically, the pattern for calling agent endpoints from a workflow is covered in our n8n AI Agent nodes review.
If you go with OpenAI and want the sandbox on your own infrastructure, DigitalOcean is one of the nine partner sandboxes for the Agents API, and its Basic Droplets start at $6 per month for a 1 GB / 1 vCPU instance ($4 per month for the 512 MB tier), billed by the second (per DigitalOcean's pricing page, fetched 2026-09-17). Spin one up and point the agent's environment at it:
# Provision a small sandbox host for the OpenAI Agents API partner integration
doctl compute droplet create agent-sandbox \
--size s-1vcpu-1gb \
--image ubuntu-24-04-x64 \
--region nyc3 \
--waitThat keeps the execution environment on hardware you audit and pay for at a flat rate, instead of metered container time you cannot see inside. It is the concrete version of the OpenAI flip condition: you get the managed harness and keep the sandbox on your own boundary.
For the framework-layer version of this decision, and the seat-versus-token economics behind running agents at all, see the enterprise agent harness decision and the rest of the Pondero enterprise desk.