Friday, 11 September 2026

OpenAI has released the Agents API in public beta, exposing the same "harness" and cloud infrastructure that runs its own Codex coding agent to outside developers [1][2]. The company says the API lets developers build and run long-running cloud agents that manage context, use tools and coordinate subagents, with the environment "fully managed by OpenAI" [1]. According to MarkTechPost, the design is built around four objects: an Agent (its model, instructions, tools and MCP servers), an optional Environment sandbox where it runs commands and loads skills, a durable Session that keeps working on a task, and the events and items that stream in and out [2].

Developers can run an agent's compute in an OpenAI-hosted sandbox, on their own infrastructure, or with no sandbox, and MarkTechPost reports OpenAI lists nine partner integrations including Cloudflare, DigitalOcean, Oracle and Vercel [2]. There is no separate fee for the API — you pay for tokens, tools and container time — but it is US-only and does not yet support Zero Data Retention [2]. One customer, Ciridae, says its evaluation score rose from 0.71 to 0.85 after switching to the API [1].

The move pushes OpenAI further into selling the plumbing around models, not just the models. For a small team, the pitch is that you no longer have to build and babysit the loop that keeps an agent running for hours — the retries, the file handling, the subagent orchestration — because OpenAI now rents it as a service [1][2].

OpenAI has temporarily stopped new sign-ups and upgrades to ChatGPT Pro, its $200-a-month top tier, saying demand for its newest model, Astra, is straining its systems [1][2]. Astra went live earlier this month, and the Pro plan — which offers the most generous access — is the one that puts the most load on OpenAI's infrastructure, so the company says it paused new Pro subscriptions while it adds capacity [1]. Existing Pro subscribers keep their access, and the API and the cheaper Go and Plus plans stay open to new users [1].

The company did not put a date on when Pro sign-ups reopen, and it did not say how many people are trying to subscribe [1][2]. Thibault Sottiaux, who leads products including Codex and ChatGPT, said demand for Astra is "unprecedented" and that OpenAI is "pulling all the levers possible to sustain" it [1]. According to Fortune, the head of product framed the freeze as "the smallest step that allows us to continue giving the broadest access possible" [2].

For anyone who was about to pay for Pro to get more Astra usage, the practical answer is that the door is shut for now. If your work depends on the highest-capacity tier, the API is still available, and it is worth checking OpenAI's status before you build a workflow around a plan you cannot currently buy [1].

Anthropic's September threat report describes what it calls illicit "distillation" campaigns that pulled Claude's step-by-step reasoning to train rival models, and it names Alibaba, Moonshot AI and DeepSeek [1][2]. According to TechCrunch, Anthropic linked nearly 200 million exchanges to distillation across five separate campaigns [1]. Distillation here means feeding a model queries, capturing its chain-of-thought answers, and using that reasoning to fine-tune smaller models [1].

The largest effort TechCrunch attributes to operators tied to Alibaba: about 151 million exchanges between May and July, peaking near 3 million a day across roughly 3,500 accounts, using a single fixed prompt to pull reasoning aimed at training the Qwen family [1]. A separate campaign linked to Moonshot AI routed some requests from the Chinese military — one example asked a model to assess whether people in surveillance footage were "behaving abnormally" — and ran through about 5,000 accounts, mostly against Claude Opus, with close to 300,000 requests in one ten-day stretch [1]. Anthropic first detailed distillation attacks and named specific labs earlier this year [1].

For anyone building on a frontier API, the read is less about espionage and more about friction: providers respond to campaigns like this by tightening identity checks, geographic limits and rate caps. Those same controls can add verification steps and slow down ordinary developers who share nothing with a state lab [1][2].

Meta's new AI agent app, Muse, has risen to No. 2 on the US App Store, up from No. 4 the day before, according to TechCrunch [1][2]. But the ranking sits on relatively small numbers: TechCrunch reports Muse has been downloaded a bit more than 83,000 times on iOS in the United States since it launched this week, and on Android it sits at No. 338 in Google Play's productivity category [1]. Muse is Meta's push into consumer agents — an assistant meant to act on a person's behalf — and is available in the US through the web, WhatsApp and mobile apps [1].

TechCrunch puts the figures in context by comparing past launches: Threads drew about 4.3 million US downloads on its first day, and even Meta's earlier standalone Meta AI app saw about 108,000 downloads at debut [1]. Dataconomy frames the same data as a modest start rather than a breakout [2].

The gap between a high chart position and a low download count is the point. App Store rank rewards a burst of installs over a short window, which a company with Meta's reach can manufacture; it says little about whether people keep using an app. For anyone deciding whether to try yet another AI assistant, the honest signal is not launch-week rank but whether users are still opening it a week or two later [1].

Nvidia CEO Jensen Huang told investors the company can grow revenue about 70% in the coming fiscal year, repeating guidance he first gave over the summer [1][2]. Speaking at the Goldman Sachs Communacopia and Technology conference, Huang said he is "confident" in the target [1]. The AI Insider notes that 70% growth on a fiscal year running near $400 billion would put next year's revenue around $680 billion [2].

Huang tied his confidence to Nvidia's position underneath the whole industry: "Nvidia runs every model," he said, listing Anthropic, OpenAI and Google among the labs that use its chips [1]. According to TechCrunch, he pointed to sales of the company's Grace Blackwell rack systems growing about 27% month over month as evidence that demand is still accelerating rather than leveling off [1].

The figure is a forecast from a vendor with every reason to sound bullish, so treat it as a claim, not a fact. But it is also a useful read on the AI economy that everyone downstream depends on: if Nvidia is right, the cost and availability of compute — and the prices of the tools built on it — stay under upward pressure for another year. If the labs buying those chips cannot turn them into revenue, the same concentration becomes the sector's biggest single risk [1][2].