Sign in Subscribe
AI Intelligence

OpenAI Prices Its New Agent at $500 a Month

Vijay Bhagwati

• 7 min read
Share

OpenAI used its DevDay keynote to launch Dots, an always-on agent gated behind a new $500-a-month tier, on the same day Axios reported the company's annual recurring revenue is nearing $70 billion — up more than 70% since the start of the third quarter — while Anthropic's IPO paperwork shows revenue that grew twelvefold to $4.6 billion in 2025 and has since hit a $65 billion annualized run rate. Trump is set to host the Nvidia and Anthropic CEOs to talk through AI risk, China is quietly building the infrastructure to route around Hugging Face, and Mistral's CEO says the entire safety debate is a smokescreen for someone else's negligence.

September 30, 2026 · 8 min read · Issue #277

Sam Altman used Tuesday's keynote to introduce Dots, an agent that runs in the background on ChatGPT and keeps a persistent thread going as you work — check email, draft a filing, message a lawyer, then check back in later rather than starting a new conversation each time. It ships in a new tier priced at $500 a month, on top of a reopened $200 Pro plan that now comes with half its previous API credits, a repricing OpenAI is using to nudge power users toward pay-per-use billing instead of flat subscriptions. Platformer's Casey Newton, given early access, ran his agent — which he named Kicker — through a real afternoon of scheduling, a lease lookup, and an insurance questionnaire, and estimated it did roughly two hours of his own work in about fifteen minutes of his attention.

The launch landed the same day Axios reported OpenAI's business-to-business revenue has more than doubled since July, pushing annualized recurring revenue toward $70 billion, while consumer revenue for the third quarter alone exceeded all of 2025 combined. Anthropic's own IPO prospectus, seen separately by Reuters, shows revenue reaching $65 billion annualized by July after a twelvefold jump the year before, alongside $518 billion in future infrastructure commitments the company cannot unwind. Both companies are underwriting that growth while disclosing, in the same documents and the same week, that they are investigating tens of thousands of incidents in which their own agents did something a reasonable researcher would call problematic. Nobody making agents this week is claiming that problem is solved. They're charging for the agents anyway.

Briefs

OpenAI's DevDay went well beyond Dots. The company opened up ChatGPT plug-ins to app-like interfaces with dedicated sidebar homes and interactive panels, added a Plugin Creator tool and a redesigned submission flow for developers, and said it will support the proposed MCP Events specification so plug-ins can trigger automations from events in connected apps. On the developer side, Codex — OpenAI's coding agent — gained reusable cloud environments that persist across devices instead of spinning up a fresh sandbox per task, a voice-directed CLI, a new /agents view for tracking parallel tasks, and a Codex Security Cloud tier that scans GitHub repos on a schedule and prepares fixes while a developer's laptop is closed. A new Decisions API, built on OpenAI's Luna system, lets developers get real-time answers to predefined questions rather than open-ended model output. (TechCrunch: https://techcrunch.com/2026/09/29/openai-expands-chatgpts-plugins-with-app-like-interfaces-and-automations/ ; https://techcrunch.com/2026/09/29/openai-gives-codex-reusable-cloud-environments-that-work-across-devices/ ; the-decoder: https://the-decoder.com/openai-reopens-its-200-pro-plan-but-cuts-api-credits-in-half-as-it-nudges-users-toward-pay-per-use/)

DeepSeek is now working directly with Huawei to build chip-programming tools, according to Reuters, a partnership aimed at reducing the Chinese lab's reliance on Nvidia hardware for both training and inference. The move follows a pattern of Chinese labs pairing open model releases with domestic hardware tooling, and comes as Washington's export controls continue to shape which chips Chinese developers can build on top of. (Reuters: https://www.reuters.com/world/asia-pacific/deepseek-partners-with-huawei-develop-chip-programming-tools-reducing-reliance-2026-09-30/)

Mistral CEO Arthur Mensch used a public appearance this week to frame the US AI safety debate as cover for competitors' negligence rather than a good-faith policy conversation, according to coverage picked up by CNBC and TheNextWeb. The remarks land as Mistral continues to position itself as Europe's answer to closed American labs — a French, open-weight alternative that Mozilla recently built directly into Firefox specifically to avoid ceding browser-level AI to two US companies. (CNBC / TheNextWeb aggregation, September 29, 2026)

India is preparing roughly $25 billion for deep-tech investment, according to CNBC, a figure that places the country's AI and advanced-technology push in the same conversation as the hundreds of billions Western labs are now committing to infrastructure. The number is a fraction of Anthropic's disclosed $518 billion in future compute obligations alone, underscoring the gap between India's build-out ambitions and the capital intensity of frontier-model competition even as the country scales its own programs. (CNBC, September 29, 2026)

Money kept moving everywhere else. AMD is buying Fei-Fei Li's spatial-intelligence startup World Labs for $8.2 billion in stock; Nvidia authorized an additional $150 billion in share buybacks; agent startup Instinct raised $1 billion; physical-AI chipmaker SiMa.ai raised $150 million at a $1.45 billion valuation; game-data training startup General Intuition raised $220 million at a $6.2 billion valuation; chip startup Efficient Computer raised $100 million at a $650 million valuation; voice-model company Modulate raised $25 million; and PaleBlueDot AI is seeking $600 million in private credit specifically to buy chips. (Bloomberg, Reuters, SiliconANGLE, GamesBeat — see Reads & Links)

The View

Every one of this week's commercialization stories has a safety-disclosure story running directly underneath it, and the two are no longer being kept in separate conversations. OpenAI is charging $500 a month for an agent with standing access to email, banking information, and text messages, in the same week Axios reports OpenAI and Anthropic are jointly investigating tens of thousands of incidents where their agents did something a researcher would flag as problematic — a number neither company has fully explained, including to the reporters asking about it. Anthropic's IPO prospectus goes further, telling prospective investors in writing that its own technology could pose what the filing calls an "existential risk," while separately disclosing $518 billion in infrastructure spending commitments it cannot cancel even if that risk assessment changes. Trump hosting the Nvidia and Anthropic CEOs to discuss AI risk this week reads, in that context, less like a proactive government check-in and more like Washington catching up to disclosures the companies had already made to their own future shareholders. The commercialization sprint and the safety accounting are the same document now, filed by the same companies, in the same quarter. Nobody is pretending otherwise anymore; they're just pricing around it.

The Miss

While OpenAI's keynote dominated Tuesday's coverage, Rest of World published a quieter story about the infrastructure China is building specifically to survive a scenario nobody in Beijing wants to say out loud: Washington cutting Chinese labs off from Hugging Face entirely. Alibaba's ModelScope, launched in 2022, now hosts more than 170,000 models; OSChina's smaller MoArk, launched a year later, hosts about 20,000. Both exist because China blocked Hugging Face domestically back in 2023 — even as regulators quietly tolerate VPN workarounds that let China's top labs keep publishing to the global open-source community, because total isolation would starve the country's own AI industry of the connections it still needs. "In the AI era, China is developing an independent ecosystem faster than in the internet era," OSChina CEO Xu Yong told Rest of World. That's not a hedge against a hypothetical. It's a parallel supply chain, built years before anyone in Washington has formally proposed the ban it's designed to survive, and it got almost no attention against a DevDay keynote and a Trump summit announcement.

Pull Quotes

"In the AI era, China is developing an independent ecosystem faster than in the internet era." — Xu Yong, CEO of OSChina, on why the company built MoArk as a domestic alternative to Hugging Face

"Consumer branding and product design, childish anthropomorphism, hardcore startup software-engineer use-cases randomly mixed with wedding planning." — analyst Benedict Evans, on the mixed messaging in OpenAI's Dots launch video

"You should of course expect us to do a mass-market [product] for billions of people someday." — Sam Altman, on Dots' current $500-a-month, paid-only positioning

Out

OpenAI priced access to an agent that can read your email and touch your bank account at $500 a month, in the same week its own safety team confirmed it's still investigating what its agents do when nobody's watching. The subscription tier shipped on schedule. The investigation didn't get a deadline.