Open AI Agents:What OpenAI's Agent Products Do, and What They Cost

On November 30, 2026, OpenAI shuts down Agent Builder, its visual tool for agent workflows. That is 17 days from today. OpenAI gave notice on June 3, 2026, on its deprecations page. The notice also covers the Evals platform and reusable prompts.

I start with that date because it tells you how this market works. OpenAI ships agent products quickly, and it retires them quickly. A guide to open AI agents is useful only if it carries the date on which the writer checked the names. I checked every name and price in this guide on November 13, 2026.

The phrase itself needs one note. People type "open AI agents" as three words. The company writes its name as one word, OpenAI. Some people mean open source agents instead. This guide covers the products of OpenAI. For the open source frameworks, read my comparison of agent orchestration frameworks.

I am Chris Machetto. I run Frenchy Digital, a software agency that started in 2016. We build agents for clients on several model providers, and OpenAI is one of them. That gives me a commercial interest, so I show the arithmetic and I give the source of each number.

This is the order. First, a map of each current product. Second, the idea that most owners get wrong. Then the ChatGPT products, the API products, the products that OpenAI retired this year, the prices and a decision table. The guide ends with what fails first, a boundary table, the limitations and three tasks for this week.

This table is the whole guide in one place. Each row is a product that exists on November 13, 2026, or one that ends this month. The status column is the part that changes most.

Two names are not in the table. OpenAI retired ChatGPT agent, the virtual computer that it launched in July 2025. The Atlas browser shut down on August 9, 2026. Both get their own rows in the section on retired products.

One more product belongs to this story. OpenAI announced Frontier, an enterprise agent platform, on February 5, 2026. The announcement page did not open for me today. A secondary report from February 13, 2026 describes limited availability for selected enterprise customers, and no public price. I give it one paragraph and no more, because I cannot verify more.

Most owners think that OpenAI sells one agent. You pay for it, you connect it to your systems, and it works. Vendors and headlines make that idea reasonable.

The real picture has two stacks, and they bill in different units. The ChatGPT stack is for your staff. You pay a seat price each month, and the agent works inside ChatGPT with the apps that an admin connected. The API stack is for your developers. You pay for tokens and tool calls, and the agent works inside software that you own.

The two stacks share models and tools, but they do not share a product. A workspace agent cannot run inside your customer app. An Agents SDK agent cannot appear in the ChatGPT window of your staff. When a vendor says "we use OpenAI agents", ask which stack.

The definition is the same on both stacks. It comes from the OpenAI guide for product teams.

Agents are systems that independently accomplish tasks on your behalf.

OpenAI, A practical guide to building agents

The guide lists three parts: a model, tools and instructions. I used that definition in my guide to AI agents for business, and it applies here. The product decides where the three parts live and who pays for them.

Three products live inside ChatGPT today: ChatGPT Work, workspace agents and Codex. Each one replaced something that came before it, so I give the history in a few lines each.

The plan prices come from the ChatGPT pricing page on November 13, 2026.

One detail on Plus matters for an agent. The page gives an estimate of 15 to 160 GPT-6.1 Sol messages in each five-hour period, and weekly limits can apply. A long Work task uses more of that allowance. Pro plans have no five-hour limit, and that is the reason that a heavy user pays $100 or more.

The API stack has three layers, and the OpenAI guide to building agents separates them the same way. The Responses API is the model call. The Agents SDK is a loop that runs in your code. The Agents API is a loop that OpenAI runs for you.

This section exists because of the hook. In the twelve months before today, OpenAI retired or scheduled the end of eight agent products or models. The table gives the dates from the deprecations page and from the reports that I name above.

The pattern is clear. OpenAI consolidates. Operator became ChatGPT agent, which became ChatGPT Work. Atlas became a browser inside the desktop app. The Assistants API became the Responses API and the Conversations API. Agent Builder becomes the Agents SDK or a workspace agent.

The notice periods differ. The Assistants API got twelve months. Agent Builder, Evals and the prompt objects got about six months. ChatGPT agent, a consumer product, got less, and users on the developer forum said that they got no notice.

That difference is useful. An API product gives you time to move. A ChatGPT product can change on a Tuesday.

The Evals shutdown has a lesson of its own. OpenAI points users to Promptfoo, and a detailed report on the shutdown says that OpenAI now describes Promptfoo as part of OpenAI. New runs there are separate from the old ones. If your tests lived in the dashboard, export them now. I explain a test set that you own in AI agent evaluation and observability.

Each price below is a list price that I read on November 13, 2026. The ChatGPT stack bills in seats and credits. The API stack bills in tokens, tool calls and container time. Keep the units apart.

Start with tokens. Assume that one support conversation uses about 20,000 input tokens and 2,000 output tokens. That is a scenario, not a measurement of your traffic. The prices come from the OpenAI pricing page.

Calculate the Sol row by hand. 20,000 input tokens at $2 for each million cost $0.04. 2,000 output tokens at $10 for each million cost $0.02. The total is six cents. On GPT-6 Luna, the same conversation costs about a third of a cent.

Tools add small amounts. The same page lists them.

Put the two together. A conversation on GPT-6.1 Sol with two web searches costs about eight cents before the search content tokens. One thousand such conversations a month cost about $80. The model is not where your money goes.

Now the ChatGPT stack. Ten people on Plus cost $200 a month. Ten people on Business cost $200 a month on an annual plan, or $250 on a monthly plan. Workspace agents then draw credits from the pool of the workspace, and the rate card gives no fixed cost for a run.

This is our own price, because I can confirm it. A custom agent from Frenchy Digital costs $5,000 plus a $5,000 setup fee, so $10,000 to start. After launch it costs $2,500 a month. Year one is $40,000 plus usage, and each later year is $30,000 plus usage. You pay the usage to OpenAI directly, and the tables above show that it is small.

Three years of a custom agent cost $100,000, which is about $2,778 a month. Compare that with ten Business seats at $200 a month. The custom agent is worth the difference in two cases. The agent must live inside your product, or it must follow rules that a workspace agent cannot hold. If your company is in New York, our team for AI agent development in New York can run that comparison with your numbers.

This is the table that I use on a discovery call. The left column is the task as the owner describes it. The middle column is the product that I would start with, not the only option.

Two rules sit behind the table. First, the OpenAI guide says: "Our general recommendation is to maximize a single agent's capabilities first." One agent with good tools beats three agents with handoffs, until the prompt holds too many branches. Second, the Anthropic guide on agents gives the same advice from the other side. It recommends "finding the simplest solution possible". Both companies sell models, and both tell you to start small.

The OpenAI guide also gives three signs that a task fits an agent at all. The signs are judgment on exceptions, rules that are too complex to maintain, and unstructured material. If a task shows none of the three, "a deterministic solution may suffice". A form and a rule cost less than any product in this article.

The project that I can show in public is the Wisdom Tooth Clinic in Miami, an oral surgery practice. We built a platform with four agents: scheduling, insurance and eligibility, intake and consent, and recall. The voice line runs on Retell. OpenAI is in the stack for one narrow job.

The case study says what that job is. Intake summaries go to OpenAI server-side, with the name of the patient stripped, under a business associate agreement. The model only summarizes. Deterministic code evaluates the warning signs in the medical history, such as anticoagulants or cardiac history. The model does not decide.

Two more details apply to any OpenAI agent. The scheduling agent books through signed, opaque offer identifiers, so the model never sees or invents a raw time slot. And the agent is not allowed to quote a price, because a Florida dental advertising statute makes an all-inclusive fee claim non-compliant.

I will not quote a result from that project. The practice opened only recently, and the case study says that the acceptance gates are not yet measured. What I can say is that the platform went from first commit to live booking in 31 days.

Now a scenario, with invented inputs. Consider an accounting firm with 15 people. Each Monday, a senior accountant spends about three hours on a report. The report pulls numbers from a practice management system and a shared drive. The firm wants an agent for that report.

• Route one: a workspace agent on ChatGPT Business. Ten seats at $20 a month on an annual plan are $200 a month. The agent runs each Monday on a schedule and posts the draft to Slack. The credits for four runs a month are not fixed, so the firm measures them.
• Route two: an Agents SDK build inside the portal of the firm. At our prices, $10,000 to start and $2,500 a month. Three years cost $100,000, or about $2,778 a month.
• The difference is about $2,578 a month, before credits. At a loaded cost of about $75 an hour for the accountant, three hours a week are about $975 a month.

I tell that firm to start with the workspace agent. Route two pays only if the report must go into a client portal, or if the data cannot go into ChatGPT at all. Both conditions are possible. Neither is the default.

Five failures occur frequently enough that you must plan for them. For each one, the table shows how you see it and what limits the damage.

The first row is the one that this guide is about. Agent Builder got a notice on June 3 and ends on November 30. The export does not convert the workflow graph. A team that kept its logic in the canvas now rebuilds it in code, in 17 days or less.

The second row has no complete solution. OWASP puts prompt injection first on its Top 10 for LLM applications. The OpenAI computer use guide says: "Computer use can affect real accounts and data." It also says: "Treat screen content as untrusted." Those two sentences come from the seller. Read them as a warning label. I explain the controls for browser agents in computer use agents and browser automation.

The third row comes from the SDK documentation itself. The sandbox guide says: "The backend determines which isolation controls apply; a session does not itself guarantee OS-level confinement." The local client on Linux adds no confinement at all. Use Docker or a hosted provider for commands that the model writes. Keep credentials outside the sandbox. The April update separates the harness from the compute layer for that reason.

The fourth row is the Agents API sentence on data residency. I repeat it because it is easy to miss on a quickstart page. United States only, and no Zero Data Retention.

The fifth row is new in 2026. Work bills against your plan allowance, and a long task uses more of it. On the API, set a hard spend limit for the project before the first run. The changelog says that OpenAI added hard spend limits for organizations and projects on July 22, 2026.

Complete this table before you switch on any product in this guide. It takes about an hour. The rows are the four places where an OpenAI agent can act.

Each product has a feature for the middle column. ChatGPT Work pauses for a sign-in or a takeover. Workspace agents let an admin require approval for a sensitive action. The TypeScript SDK has a human-in-the-loop guide and guardrails. The Agents API has automatic review, where a model approves tool requests. I would still keep a person on each write.

The OpenAI guide names the two triggers for a person. The first is a failure threshold, such as an agent that fails to understand a customer after several attempts. The second is a high-risk action. The examples in the guide are "canceling user orders, authorizing large refunds, or making payments." Put those three in the right column on day one.

These are the signs that tell me not to buy, from a product or from a builder, and that includes us.

• The vendor names a product that no longer exists. OpenAI retired Operator, ChatGPT agent and the Assistants API, and Agent Builder ends this month. A proposal dated this month that builds on them is out of date.
• The vendor cannot say which stack. Ask: does the agent run inside ChatGPT or inside our software? If the answer is unclear, the price is unclear too.
• An accuracy or deflection rate with no method. In this market, the seller publishes each such figure. Ask for the test set.
• No test set at all. The Evals platform ends on November 30. Ask where the tests live now.
• A quote before the vendor asks what the agent must read and write. The integration is the cost, not the model.
• Data terms in one sentence or less. For the Agents API, ask about residency and retention. For a workspace agent, ask which connectors the admin allowed.
• Security badges with no report. Ask for the document, not the logo.
• A contract with no exit: no source code, no export, no end date.

On the seventh point, this is our own position. Frenchy Digital signs business associate agreements with healthcare clients. We hold no SOC 2 or ISO 27001 attestation. If your procurement process requires one, a vendor that has it will pass and we will not.

This is what I did not verify, or decided not to use as evidence.

• Every page on openai.com and help.openai.com answered 403 to my requests on November 13, 2026. I cite only pages that loaded. Those are the developer documentation, the ChatGPT documentation, the SDK documentation, the GitHub repository, the PDF guide and the reports that I name.
• The status of ChatGPT agent comes from two secondary sources. One is a tool directory that read the help page on August 30, 2026. The other is a support reply on the developer forum. I did not read the help page myself.
• The Frontier facts come from one secondary report of February 13, 2026. I use only the launch date, the limited availability and the absence of a public price.
• The general availability date of workspace agents and the current credit rates are on OpenAI help pages that I could not open. I use the July 6, 2026 pricing date from a report that cites the rate card.
• The nine sandbox partners of the Agents API come from a trade report. The overview page that I read does not name partners.
• Each price is a list price on November 13, 2026. OpenAI changes prices, and the promotional prices on the page have end dates.
• The token scenario and the accounting firm scenario are arithmetic on assumed inputs, not measurements.
• The OWASP site shows its 2025 list as current, and some security firms describe a 2026 edition with a different order. I state only that prompt injection is first, which is true in both.
• I did not verify the first release dates of the Agents SDK and the Responses API today, so I do not print them.

1. 1.Open the OpenAI deprecations page. Search it for each product name in your current proposals and contracts. Write down each shutdown date that you find.
2. 2.Decide who the first agent works for: your staff, your customers or your product. That answer selects the stack, and the stack selects the price unit.
3. 3.If the answer is your staff, switch on ChatGPT Work or a workspace agent for one task, in a plan that you pay for now. Measure ten runs before you decide anything else.

If you complete those three tasks and want a second opinion on the stack, book a discovery call with us. If the ChatGPT stack is enough, I will tell you.

Start with the deprecations page.