The Debrief·No. 26·September 15, 2026
OpenAI wants to run your agents
The whole agent loop is now one API call with no platform fee. Two lines in the docs decide whether you can actually use it.
Welcome to the Debrief, the Tuesday half of The AI Debrief. This is what happened: the few AI moves that change what you build, and the ones you can safely skip. Friday is the Toolkit, where we hand you something to use.
Here's what we're getting into this week. OpenAI turned the harness behind Codex into a hosted product and priced it at zero. DeepSeek put a 1M-context model on the API at fifteen cents per million tokens with open weights. Anthropic named the labs it says were copying its models at industrial scale. And Cursor shipped an agent that is meant to run for months.
In this email » The Agents API: nine sandbox hosts, no fee, and the retention clause that will stop half of you » DeepSeek V4.1-Flash at $0.15 per million, and the routing reversal to check before you migrate » 200 million exchanges: Anthropic names Alibaba, Moonshot and DeepSeek » Cursor Projects, and the 99% benchmark score you can ignore |
| 🤖 The harness is free. The data is not yours. |
OpenAI put the whole agent loop behind one API call
On September 10 OpenAI opened the Agents API in public beta. You send a task, a model, a set of tools and an environment, and OpenAI runs the loop: sessions that persist, context that compacts itself as it fills, subagents that fan out with their own context, MCP servers, custom functions and built-in web search. The default model is gpt-6-astra. The sandbox where the code runs can be OpenAI's own or one of nine partners: Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, Vercel and Blaxel.
The price of the orchestration layer is zero. OpenAI's exact words are that there are no additional fees, you pay for the tokens and tools your agents use. Hosted sandboxes bill at their standard container rates. This is the same harness that runs Codex, and until last week you had to build your own version of it or pay a startup for one. A lot of those startups had a worse week than it looked.
Platform fee $0 | Sandbox hosts 9 + OpenAI | Zero Data Retention Not offered |
Now the two lines. The developer docs say the Agents API currently supports data residency only in the United States and does not support Zero Data Retention. Then, one paragraph later, the part I would have missed: picking a self-hosted sandbox does not change that. Your code can run on your own Modal account and the session data is still held under OpenAI's standard retention. If you sell into anyone with a data-processing agreement, a European customer, or a security questionnaire that asks about ZDR, this beta is a demo, not a deployment. That is a beta limitation and it may lift. It is also the exact clause enterprise buyers check first, and OpenAI shipped without it.
Same day, two smaller things that belong in the same picture. OpenAI put GPT-Live-1, its full-duplex voice model, into the API at a flat five cents a minute for the voice layer, with the text model underneath billed separately. Every benchmark on that page is OpenAI's own or a launch partner's; the price is the fact. And OpenAI paused new $200-a-month Pro sign-ups, citing demand for Astra, with no end date. Existing Pro, Plus and the API are unaffected. If your team is about to expand Pro seats, that is a queue now.
I think the Agents API is the most important developer release of the year so far, and I am not going to build on it this month. Both things are true. The harness being a commodity means the value you add is the tools, the permissions and the data you connect, which is where it should have been anyway. The retention gap means the version worth building on is the one that ships when the beta label comes off.
Why you care If you maintain your own agent loop, this is the week to write down what it does that OpenAI's does not. If the answer is nothing, plan to migrate when ZDR arrives and stop investing in the loop now. If you run agents on customer data, add "ZDR available" to the checklist you run on every agent platform, because this week proved the biggest vendor can ship without it. |
| ⚡ Worth your time |
DeepSeek V4.1-Flash: 1M context, open weights, fifteen cents a million.
On September 10 DeepSeek released V4.1-Flash: a 552B mixture-of-experts model with 8B parameters active on input and 16B on output, MIT-licensed weights on Hugging Face, and a KV cache DeepSeek says needs a quarter of the memory of the previous generation. The pricing page has it at $0.30 per million uncached input and $1.20 output at peak, half that off-peak, with 1M context and 384K max output. The API name is deepseek-flash; the old deepseek-v4-flash is retired and temporarily aliased. Two cautions. The benchmarks are DeepSeek's, and the one independent-ish read I trust is the gap on Terminal-Bench, where it scores 30 against Opus 5's 43, so it is a cheap model for high-volume work, not a frontier replacement. And the launch post said all V4-Pro traffic would be routed to Flash from September 14, while the pricing page now says V4-Pro continues after September 14 with billing unchanged. The later statement is the pricing page. If you call V4-Pro, check which one is true before your next bill, not after.
Anthropic says 200 million of its responses went to three Chinese labs.
Last week's US government advisory on distillation named the labs but not the numbers. On September 10 Anthropic's threat report put figures on it: five campaigns, nearly 200 million exchanges, the largest attributed to Alibaba at 151 million exchanges between May and July across 3,500 accounts, peaking near three million a day, and one attributed to Moonshot at roughly 300,000 requests in ten days through 5,000 accounts. All of it is Anthropic's attribution; none of the named companies has responded that I can find, and I would hold the per-company numbers loosely. What is not in the report is any change to the API, rate limits or verification for you. Read it as the evidence file for the policy fight, and as a reminder that "cheap and nearly as good" sometimes has a supply chain.
Cursor Projects: an agent that runs for months and watches your Slack.
On September 10 Cursor started rolling out Projects in beta: a cloud-hosted coordinator thread for a feature, a migration or a whole app, which keeps its own context files, delegates to subagents and can be triggered by a Slack channel, a schedule or pull-request activity instead of by you. Cursor's number is that users who primarily work in Projects merge six times as many PRs, which is a self-selected group measuring itself. No pricing and no model list are on the page, and given that Cursor loses OpenAI models on November 12, "which model is the coordinator" is the question to ask before you move real work into it. The pattern is the real story: OpenAI shipped hosted subagents, Cursor shipped a coordinator, and both in the same week. The unit of work is no longer a chat.
| 🙅 Skip it |
GPT-6 Astra "beats Portal" and scores 99% on ARC-AGI-3 A single blog post this weekend reported that Astra, in a custom harness, finished the 2007 game Portal in just under 24 hours and about $570 of tokens, cleared 80 of 154 Baba Is You levels, and posted 99 percent on ARC-AGI-3. It is a fun read. It is also one team, one unreviewed harness, and a benchmark nobody else has run yet. There is nothing to buy, call or reproduce, and the score will look different once the benchmark authors weigh in. File next to last week's Navier-Stokes proof: interesting, unverified, not a build decision. |
The single thing to do this week: open the docs for every agent platform you pay for and search the page for "retention". Honest worst case: ten minutes confirming you are covered. Likely case: you find one vendor that never said, and now you know which question to send them.
Every play this newsletter has shipped lives with the people testing them in the free AI Academy.
| Join the free AI Academy → |
Free, no card. The tools and people behind this issue live here.
That's the Debrief.
See you Friday for the Toolkit.
From Drew and The AI Debrief team
P.S. Are you building your own agent loop, or waiting for someone else's? Hit reply and tell us which. We read every one.
P.P.S. New here, or skipped one? Every past edition is in the archive.