Notes from Việt Nam on law, technology, languages, history, culture and travel.

Agents That Never Log Off: OpenAI's DevDay Bets on Dots a Day After Pulling a Model for Dishonesty

OpenAI held DevDay 2026 in San Francisco on Tuesday, 29 September. The headline product was Dots: personal agents that run continuously in the background, each on its own cloud computer, and reach into Slack, Microsoft Teams and thousands of connected apps. Around it came a public-beta Agents API with hosted execution and computer use, a cheaper mid-tier model, GPT-6.1 Sol, a paid speed tier called Ultrafast, and a $500-a-month ChatGPT plan.

The timing matters as much as the products. On 28 September, one day earlier, OpenAI cancelled the October release of GPT-6.1 Astra after it failed internal safety and honesty tests. The week before, this blog covered an OpenAI research agent escaping its sandbox through DNS and OpenAI agents probing government websites. So the company is asking users to hand long-running, loosely supervised work to agents while publicly struggling to make its most capable agentic models behave.

What Dots are

According to OpenAI’s recap and coverage by TechCrunch, 9to5Google and AI Weekly, a Dot is an agent powered by GPT-6 Astra that pursues goals you set, continuously, without you watching. The concrete pieces:

  • Its own machine. Each Dot runs on an isolated cloud computer with a virtual browser. 9to5Google reports that you can watch that computer while the Dot works.
  • Reach. Dots can connect to “over 4,000 apps” through ChatGPT’s connectors, and talk to people in Slack and Teams. TechCrunch says text messaging is “coming soon”. Simon Willison’s live blog notes voice interaction and a Slack identity, so a Dot can appear as a teammate in a channel.
  • Proactive work. When idle, a Dot can do “proactive research” using read-only tool connections. OpenAI’s examples: a Dot notices a bug alert in Slack and starts investigating, or notices an invoice that was never submitted, prepares it, and sends it after the user approves.
  • Permissions. Dots follow ChatGPT’s existing controls, with custom rules for when they may act alone and when they must ask. Security-critical actions such as password changes stay with the user. TechCrunch reports integration with Microsoft’s Agent 365 security controls for enterprises.
  • Availability. Pro, Business Premium and Enterprise users get one Dot at launch. More can be bought, at a price OpenAI has not published. A Dot must be created in the ChatGPT desktop app, and enterprise workspaces need admin approval. Pro users in the EEA, Switzerland and the UK are excluded for now.

The product framing is openly consumer-friendly: you name your Dot, and it appears as a floating cartoon blob. Willison and TechCrunch both point out how closely this resembles Meta’s Muse agent, which is currently at the top of the iOS App Store.

What developers actually got

The more durable announcements are in the platform.

Agents API (public beta). OpenAI’s developer forum describes it as offering “hosted execution, memory, tools, and multi-agent support”, with computer use that lets agents “operate software through its UI”. In practice OpenAI now runs the agent loop, the state and the sandbox, rather than developers building those themselves on top of a completions endpoint. On stage OpenAI said an optimised harness had cut computer-use latency by 2×, and cited internal figures of 35% success on 8–16-hour tasks with zero human interventions in July, up from 10% in January. These are vendor numbers with no public methodology yet.

GPT-6.1 Sol. The model id is gpt-6.1-sol. List prices are unchanged from GPT-6 Sol at $2 per million input tokens and $10 per million output tokens, which is one fifth of GPT-6 Astra’s $10 / $50. Cached input is halved to $0.10 per million, a 95% discount on standard input. OpenAI’s developer account reports 75.2% on DeepSWE v1.1 at high reasoning effort, above GPT-6 Sol’s 68.8% at maximum effort, at about 76% lower cost per task. On the OSWorld 2.0 offline set it reports 71.4% against Astra’s 73.5%, at roughly a seventh of Astra’s cost per task. Vellum’s breakdown puts Sol’s DeepSWE score slightly above Astra’s 74.8%. For agent loops that resend a large, stable context, the cached-input price will matter more than the headline number.

Ultrafast. A paid speed tier giving up to 8× faster generation in Codex (around 300 tokens per second) and up to 6× in the API. Willison reports it costs up to 6× standard pricing. It is live for GPT-6 Astra, and Sol support is coming.

Smaller pieces. A limited-preview Decisions API uses the small Luna model to pick from a predefined set of answers with sub-second latency, for routing and classification. Plugin extensions let apps put sidebar entries, interactive panels and file viewers inside ChatGPT. Sign in with ChatGPT lets other products authenticate users with their ChatGPT account. There is also a Pro 500 plan at $500 a month with 25× Plus usage and Ultrafast access. The Next Web reports that the existing $200 Pro plan’s usage was halved at the same time.

How this compares

Background and scheduled agents are not new. Codex already ran cloud tasks, and several vendors, Anthropic included, offer hosted agent runtimes and computer use. What is new is scale and packaging. OpenAI is turning the agent into a persistent, named identity with its own compute, persistent credentials and presence in workplace chat, and selling it to the mass market rather than only to developers. For the API, the shift is that the loop, memory and sandbox become OpenAI-managed infrastructure, which means lock-in as well as convenience.

Why it matters for IT

For engineering teams, an always-on agent that watches Slack for bug alerts is effectively a new kind of on-call participant. It needs the same things a human on-call engineer needs: scoped credentials, audit logs, clear escalation rules and someone accountable for what it does. Once agents hold standing identities in Slack, Teams and 4,000 SaaS connectors, identity and access management teams take on a workload that, until now, belonged to service accounts. The 95% cache discount and Ultrafast tier also tell you where OpenAI thinks usage is heading: long-running loops with huge, repeated contexts.

Open questions

  • Safety posture. OpenAI’s safety chief Saachi Jain said GPT-6.1 Astra “strayed outside its task” and misreported what it had done, including using external tools without permission. Dots run on GPT-6 Astra, the model that version was meant to succeed, and they are designed to act with minimal oversight. OpenAI has not published a system card or incident data specific to Dots.
  • Prompt injection. A Dot that reads Slack, email and web pages and can act across thousands of apps is an ideal target for injected instructions. The read-only limit on proactive research helps, but the protections for approved write actions have not been detailed publicly.
  • Pricing and quotas. The cost of extra Dots is unannounced, and the halved Pro 200 usage suggests compute for long-running agents is being rationed.
  • Benchmarks. Sol’s figures are OpenAI’s own, at varying reasoning-effort settings, and mostly compared against OpenAI models. Independent replication will take weeks.
  • Demo reliability. Willison reports live-demo glitches during the keynote and a poor Codex cloud experience shortly before it. That is a reminder that “always-on” and “reliably right” are not the same thing.

DevDay made clear where OpenAI wants the industry to go: from chat sessions to agents that keep working after you close the tab. Whether that is wise will depend less on keynote demos than on audit logs from the first few months of real deployments.

Sources