OpenAI's Public Agent API: What the September 2026 Beta Costs and How It Compares

On September 10, 2026, OpenAI put its Codex coding-agent harness behind a single public agent API, in beta, with no platform fee beyond model tokens and the tools an agent actually calls.

The pitch is narrow but specific: building a reliable agent requires a harness that manages context, calls tools efficiently, coordinates subagents, and keeps state alive across sessions that can run for days — infrastructure most teams currently rebuild from scratch or buy from a framework vendor. OpenAI is now renting out the version it already built for its own products.

As of October 30, 2026 the Agents API is still in beta, the $0 platform fee still holds, and the price of the model underneath has fallen: GPT-6.1 Terra, released October 26, costs $1.00 per million input tokens and $5.00 per million output tokens. This update also corrects one claim from the original version of this piece, about Anthropic.

Public Agent API: What OpenAI Actually Shipped on September 10

OpenAI's public agent API is a managed service, in beta since September 10, 2026, that hands any developer the same session-management, sandboxing, and subagent-orchestration harness that runs Codex, OpenAI's own coding agent, behind a single API call. Instead of writing the plumbing that keeps an agent's context, files, and tool calls consistent across a multi-hour or multi-day task, a developer calls the API and lets OpenAI run that infrastructure.

Is the OpenAI Agents API still in beta?

Yes. OpenAI's changelog records the September 10 release as a public beta, and the current Agents API overview still shows an OpenAI-Beta: agents=v1 header in its request example. Beta status matters for planning: behavior and billing can change between releases, and the October 9 change that began billing hosted-sandbox compute during active turns is an example of a cost-relevant change that landed after launch.

What Has Changed Since the September 10 Launch?

Six dated entries in OpenAI's API changelog touch the Agents API or its neighbors between September 10 and October 30, 2026:

Source: OpenAI API changelog, checked October 30, 2026.

The pattern is a platform deepening around hosted execution: more tools inside the sandbox, a way to hide start-up latency, and faster inference. The one change that touches the bill directly is the October 9 billing entry, which is why the sandbox math below matters more than it did at launch.

The Four Objects the API Is Built From

Everything in the Agents API reduces to four pieces, according to OpenAI's own launch post and MarkTechPost's technical breakdown of the release:

Inside a live session, the API can run code, edit files, search the web, apply predefined skills, produce artifacts, and split work across subagents under a configurable concurrency limit — the same capability set Codex uses internally, according to MarkTechPost's reporting on the launch.

The Agent object's ability to connect to MCP servers is a quieter part of the announcement but arguably a load-bearing one. Model Context Protocol, the open standard Anthropic introduced in late 2024 for connecting AI systems to external tools and data, is now the common plumbing underneath both of OpenAI's agent products — the Agents SDK and the new Agents API — as well as Anthropic's own agent tooling. Cross-vendor tool compatibility is not the industry default, so a developer's MCP server built for one lab's agent product has a real chance of working unmodified against another's, which lowers the switching cost the rest of this piece argues OpenAI is otherwise trying to raise.

Where the Compute Actually Runs: 9 Partner Sandboxes

Developers choose where an agent's Environment executes: an OpenAI-managed sandbox, their own infrastructure, or one of nine named partner sandboxes — Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel. Each partner bills its own compute rates directly; OpenAI adds no markup on top.

That partner list is notable for one name in particular: Oracle is simultaneously a named sandbox partner and, according to MarkTechPost's reporting, one of OpenAI's own largest compute suppliers under the Stargate buildout — which puts OpenAI in the position of routing developer workloads to infrastructure it also buys from at massive scale.

The choice matters beyond raw price. A regulated enterprise that cannot send proprietary code or customer data into an OpenAI-controlled environment can pick a self-hosted sandbox, or a specific partner whose data-residency and compliance posture it already trusts, without losing the session-management and orchestration layer OpenAI still runs on top. That flexibility is the part of the pitch aimed less at indie developers and more at the enterprise buyers who were the actual holdout on adopting hosted agent infrastructure to begin with.

Does the Agents API support Zero Data Retention?

No, and the sandbox choice does not change that. OpenAI's Agents API overview says the API "currently supports data residency only in the United States and does not support Zero Data Retention (ZDR)," and that "choosing a self-hosted sandbox does not make the Agents API ZDR-eligible." A regulated team can still keep code execution on its own infrastructure, but the session itself runs on OpenAI's side, so anyone who needs ZDR or non-US residency has to use a different path.

How Much Does an OpenAI-Hosted Sandbox Cost?

OpenAI's docs say hosted sandboxes "use standard container rates," with model usage billed separately at the selected model's API rates. OpenAI-hosted sandboxes come in three sizes, and OpenAI's pricing page lists container rates per 20-minute session by memory size. The memory sizes line up with the 1 GB, 4 GB and 16 GB rows, so those rows are the best published guide to hosted-sandbox cost; confirm with a test session before budgeting, because OpenAI does not spell out the size-to-row mapping.

Source: OpenAI, OpenAI-hosted sandboxes and API pricing, checked October 30, 2026. The pricing page also lists a 64 GB row at $1.92.

What Do Agent Tokens Cost After GPT-6.1 Terra?

The Agents API charges the selected model's standard rates, so the October 26 release of GPT-6.1 Terra changes what an agent costs. To make the gap concrete, take one illustrative workload of 1,000,000 input tokens and 100,000 output tokens (an assumption for the math, not an OpenAI figure) at standard short-context rates:

Source: OpenAI API pricing, checked October 30, 2026; workload sizes are an illustrative assumption. Cached input, Batch and Fast mode change these totals; see our OpenAI API pricing guide.

What Sandbox Time Actually Costs, Provider by Provider

OpenAI charges nothing extra for the Agents API itself, but the sandbox a developer picks still has its own bill. Here's what the researched partner and competitor pricing looks like as of September 2026 — every figure below comes from each vendor's own published pricing or vendor-pricing analysis, not from OpenAI.

Source: OpenAI Agents API announcement (Sept 2026); E2B published pricing, via Morph's E2B pricing breakdown; TrueFoundry's LangGraph pricing analysis; Upstash's agent sandbox provider comparison, 2026.

As of October 2026, LangChain has simplified LangSmith Deployment's pricing to two self-serve tiers: the $99/month Professional plan has been retired, leaving Plus at $39/seat/month (one free small deployment included) plus usage, and a custom-priced Enterprise tier — see LangChain's pricing page.

Agents API vs Agents SDK: Two Different Products, Same Name Confusion

The single most common mistake in early coverage of this launch is conflating the new Agents API with the existing OpenAI Agents SDK — they are genuinely different products built for different jobs. The Agents SDK is a free, MIT-licensed Python and TypeScript framework that launched in March 2025 as the production successor to OpenAI's experimental Swarm project. Its Runner executes the agent loop — tool calls, handoffs, guardrails — inside your own application and your own infrastructure; OpenAI updated it again in April 2026 with sandboxing support and provider-agnostic access to more than 100 LLMs beyond OpenAI's own models.

The Agents API is the opposite deployment model: OpenAI runs the session, the context compaction, and (optionally) the sandbox for you, and your application just calls the API and consumes the event stream. One is a library you embed and control; the other is infrastructure you rent. A team that wants full control over deployment, storage, and approval logic still reaches for the SDK. A team that wants to skip building session infrastructure altogether is who the API is for.

How the Field Compares: LangGraph, Vercel AI SDK, and Anthropic's Agent Products

OpenAI isn't launching into an empty market. Several other paths already exist for a team that wants to ship an agent without hand-rolling infrastructure, and each takes a different bet on where the line between "framework" and "hosted service" should sit.

Source: OpenAI product documentation; TrueFoundry LangGraph pricing breakdown; Vercel AI SDK 6/7 release notes; Enterprise DNA reporting on Anthropic's August 2026 GA, 2026.

Anthropic works on two layers. It has been generally-availabling individual capabilities piecemeal — computer use and browser use moved out of beta on August 19, 2026, and self-hosted sandboxes with MCP tunnels shipped on May 19, 2026, letting enterprise teams keep execution on their own infrastructure. It also offers Claude Managed Agents, a hosted harness that is the closest counterpart to OpenAI's Agents API. An earlier version of this piece said no single Anthropic product mapped onto OpenAI's bundle; that was wrong, and the next section corrects it.

What is Claude Managed Agents, and how does it compare?

Anthropic's documentation describes Claude Managed Agents around the same four concepts as OpenAI's API: an agent, an environment, a session and events. Environments can be an Anthropic-managed cloud sandbox or a self-hosted sandbox. It is also in beta, behind the managed-agents-2026-04-01 header, and its reference page lists it as not eligible for Zero Data Retention. Anthropic's release notes add that on October 9, 2026 an agent can be given dynamic workflows (beta), where it writes a program that runs many agents in phases for work such as reviewing hundreds of documents.

Source: Anthropic, Managed Agents overview, reference, pricing and release notes; OpenAI documentation as cited above; all checked October 30, 2026. Token rates are list prices, not a quality comparison.

Anthropic's pricing page says Managed Agents bills model tokens at standard rates plus session runtime of $0.08 per session-hour, counted only while the session is in the running state, plus $10 per 1,000 web searches, with no batch discount. Its worked example is a one-hour Claude Opus 5.5 session with 50,000 input and 15,000 output tokens, which comes to $0.58. Anthropic also limits each organization to 300 create requests and 1,200 read requests per minute on Managed Agents endpoints, per its reference page. The practical difference: OpenAI folds harness cost into token and container rates, while Anthropic prices the harness as an explicit hourly line.

What Is the Decisions API, and Is It Part of the Agents API?

No, it is a separate product that shipped alongside it. OpenAI released the Decisions API in beta on October 6, 2026 with gpt-6-luna, and its current guide says it is now generally available. The guide says gpt-6-luna is the only supported model, lets a request ask up to 400 questions (up from 200 during the beta), charges $0.10 per million uncached input tokens with cached input at standard gpt-6-luna caching rates, and does not charge output tokens. OpenAI says it returns typed answers about 10x faster than the Responses API. For agent builders the relevance is routing: classification-style steps can run on a cheap structured endpoint while the Agents API handles the long-running work. Source: OpenAI Decisions API guide, October 30, 2026.

Where the headline misses

The "no extra fee" framing that dominated launch-day coverage undersells what this move actually does. OpenAI isn't giving away infrastructure out of generosity — it's monetizing the layer underneath it more aggressively. Every workload that runs through the Agents API, including the tool calls, subagent turns, and context-compaction passes the harness generates automatically, consumes model tokens that OpenAI bills for. A framework that used to sit on top of a model API and take its own margin is now replaced by infrastructure that routes more spend directly to OpenAI's own models. That's a funnel into token revenue, not a free product.

There's also a real lock-in question the "$0 platform fee" pricing obscures. Once a team's session state, context-compaction logic, and subagent orchestration all live inside OpenAI's hosted harness rather than in code they control, switching to Claude, Gemini, or an open model later means rebuilding that plumbing from scratch — the exact cost the API is designed to make a developer forget about until they try to leave.

What This Means for Agent-Framework Startups

The competitive read that surfaced fastest after launch: OpenAI is competing directly with the field of agent-orchestration startups — LangChain, CrewAI, and a wave of smaller session-management layers — by commoditizing the exact harness those companies have been charging for. A framework whose entire pitch is "we handle session persistence and sandboxing so you don't have to" now has a much harder sell when the model vendor those agents already call gives that away as part of the API bill.

Early developer reaction on Hacker News mixed genuine interest with the same lock-in concern raised above: one recurring comment noted that OpenAI and Anthropic remain "stuck using their own proprietary frontier models," which for some multi-model teams outweighs the convenience of skipping the infrastructure build. This piece cites no adoption or usage numbers for the Agents API because none turned up in the sources reviewed; it has been in public beta for about seven weeks as of October 30, 2026. Any claim about how fast it's being adopted would be a guess, not a sourced fact.

For anyone evaluating agent-orchestration startups as an investment or a build-vs-buy decision, this launch is the cleanest test case available: how much of a portfolio company's retention was genuine workflow lock-in, specific integrations, vertical tuning, versus simply being the easiest way to get durable sessions and sandboxing before OpenAI shipped this API for free. Genuine workflow lock-in survives a launch like this one. Retention built only on convenience, now priced at a $0 platform fee, does not.