What is the OpenAI Agents API? Why “No Extra Fee” Is the Most Expensive Sentence in the Launch

OpenAI put the Codex harness behind one API call and charged nothing for the harness itself. I went and read the pricing page anyway. The sandbox everyone is panicking about is the cheap part.

On Wednesday, September 10, 2026, OpenAI pushed the Agents API into public beta, and my timeline split in about nine minutes.

One half was celebrating the death of the orchestration layer. The other half was quoting container rates in horror and calling it a trap.

Both halves are arguing about the wrong number. I pulled the actual rates, did the arithmetic, and the expensive part is not where anybody is pointing.

Author's Note to Readers:

As a non-native English speaker, I utilize AI language tools to assist in polishing my writing, ensuring clarity and professionalism in my communication. My blog is mostly dedicated to exploring topics within the artificial intelligence sphere, including innovative technologies and AI-generated imagery. To enrich the discussion and provide tangible examples, AI-generated images and videos are occasionally incorporated and are always clearly labeled as such in their captions.

The integration of AI assists not only in language refinement but also in the thematic presentation of content that is central to AI's impact and applications. This disclosure is made to maintain transparency with my readers and to align with platform guidelines regarding AI-assisted content. My commitment is to provide thoughtful and informative content that deepens your understanding of AI technologies, while clearly indicating any AI-generated or AI-assisted elements within my posts. Transparency forever.

We Have All Written This Loop Twice

You already know this loop by heart.

Call the model, parse the tool calls, run them, append the results to history, then check whether you've blown the context window.

Summarize when you have. Retry when a tool times out. Persist all of it, because a crash at hour four should never cost you hour one.

I have written that loop three separate times, and it was never the interesting part of any of those projects.

It's plumbing. It's also where most of my agent experiments quietly died, not on reasoning but on a session that lost its state overnight.

So What Is the Agents API?

It's OpenAI selling the Codex harness directly. That means the same context management, tool routing and subagent coordination that runs their own coding agent.

I POST to https://api.openai.com/v1/agents/sessions with the header OpenAI-Beta: agents=v1, and that is the whole integration.

Key pieces:

• Agent — the model, the instructions, the tools, and any MCP servers.
• Environment — a sandbox where it can touch files and run code. OpenAI-hosted, self-hosted via codex exec-server, a partner (Cloudflare, Modal, E2B, Daytona, DigitalOcean, Oracle, Runloop, Vercel, Blaxel), or none at all.
• Session — persistent state that survives across turns.
• Events and items — what goes in, what comes out.

Three built-in tools ship with it: programmatic_tool_calling, mcp, and web_search. Multi-agent mode takes a max_concurrent_subagents setting, which the docs demo at 4.

The headline feature is that sessions compact themselves. As a session nears its context limit the API summarizes the earlier turns automatically, so it can keep working for hours or days without you writing a single line of summarization logic.

Hold on to that sentence. We're coming back to it in a minute, and it changes shape when we do.

Example 1: The Container Everyone Is Scared Of

Here's what the hosted sandbox costs, per 20-minute session per container:

• 1 GB — $0.03
• 4 GB — $0.12
• 16 GB — $0.48
• 64 GB — $1.92

Billed by the minute, five-minute minimum.

So run a 4 GB box for a full eight-hour workday. That's 24 blocks of 20 minutes at twelve cents each, which comes to $2.88.

Now go absurd and take the 64 GB box for those same eight hours: $46.08. That's the ceiling, on a machine most agents will never need.

The container is not your problem. It costs less than my lunch.

Example 2: The Number Nobody Is Quoting

gpt-6-astra is what the docs use in every example. Standard rates, per million tokens:

• Short context: $10 in, $50 out
• Long context: $20 in, $75 out
• Fast mode: $20 in, $100 out — doubling again on long context

Now price one ordinary turn, where your agent re-reads 60k tokens of working context and writes 3k back.

That's $0.60 in and $0.15 out. Seventy-five cents.

Run it 200 times across an afternoon — a modest number for something built to work for days — and you're at $150, sitting on a sandbox that cost under three dollars.

Example 3: Compaction Is a Billing Feature

Look at those two rate rows again. Long context doubles the input price and takes output from $50 to $75.

Price that same 200-turn session at long-context rates instead and it becomes $1.20 in and about $0.23 out per turn, which lands you at roughly $285.

That's not a feature difference. That's identical work, on the wrong side of a line.

Which reframes compaction entirely. OpenAI presents it as a convenience — no summarization logic to write. It is also the only thing standing between a long-running session and the higher tier.

I don't think that's cynical of them. I think it's quietly the most valuable thing in the launch, and it got filed under "quality of life" instead of "cost control."

Subagents cut the other way, though. Four concurrent subagents, each holding its own context, is four contexts being billed. Parallelism buys wall-clock time, never a discount.

The Numbers OpenAI Put On Stage

Three customers got quoted in the launch:

• Ciridae — eval score from 0.71 to 0.85, and a 4× latency reduction.
• SafetyKit — 60% lower cost per case.
• Hypha — 86% fewer failed responses.

Real companies, real improvements, and I have no reason at all to doubt any of them.

But all three are design partners who were building against this months before either of us could open the docs. SafetyKit's 60% is measured against SafetyKit's own old pipeline — not against yours, and not against whatever harness you'd throw together this week. No baselines were published.

Read them as proof the thing works. Don't read them as a forecast of your invoice.

Where It Falls Down

No Zero Data Retention. That single line disqualifies a pile of regulated work before the pricing conversation even starts.

US-only data residency. If you're running anything in the EU, that's where your evaluation ends.

Locked to OpenAI's frontier models. You can't swap something cheap in for the boring subagent turns. Model choice is the biggest cost lever in agent work, and this API quietly takes it off your table.

It's a stateful OpenAI API. Developers on HN said it plainly: they got burned when the Assistants API was retired, and they are not handing their state over a second time. That isn't paranoia, it's pattern recognition.

It's beta. The header says so.

So Is "No Extra Fee" True?

Yes. Completely. You pay for tokens, tools and container time, with nothing stacked on top.

It's also the least interesting true sentence in the entire announcement.

OpenAI isn't charging you for the harness. It's charging you for what a good harness makes easy, which is running longer, reading more context, and spawning more agents. Those were the expensive things before Wednesday, and they are still the expensive things today.

Take It or Build It?

I'd take it when the bottleneck is the loop itself. If you've rewritten that while-loop twice and resented it both times, this is a genuinely good afternoon and I would use it without hesitating.

I'd skip it when a working harness already exists, when ZDR or EU residency is non-negotiable, or when swapping models is how the budget actually gets controlled.

And either way: read the container rates, then read the token rates, in that order. One of the two will surprise you, and I promise it isn't the one everybody is shouting about.

P.S. Follow me on X ( @lvntblsn ) if you want to watch me run an eight-hour agent session purely to see what the bill looks like. I will absolutely post the screenshot, good or bad.