OpenAI Shipped Dots. Practitioners Shipped Skepticism.

OpenAI DevDay 2026 had one headline: Dots, always-on agents powered by GPT-6 Astra, each running on its own cloud computer with persistent memory and access to 4,000+ app plugins. Sam Altman called them "a new way to use AI that works 24/7 for you." Hacker News, with 750 points and 629 comments, called them something else entirely — "a dumbed down reskin of Codex/ChatGPT Work but with the power-user features removed."

📖 Read the full version with charts and embedded sources on AgentConn →

Both sides are telling you something real. The gap between them is where the operator signal lives.

View original post on X →

Here is the thesis that most DevDay coverage missed: Dots is not OpenAI's best product announcement this week. It is their best distribution announcement. The buried number — 1.2 billion ChatGPT weekly active users — is the real story. Dots is the mechanism that turns passive ChatGPT users into always-on agent consumers, and the 4,000-plugin ecosystem is the integration moat that makes switching expensive. The actual productivity tooling for builders — Codex Cloud, Ultrafast, GPT-6.1 Sol at one-fifth of Astra's price — got second billing but deserves first attention.

If you build or operate agents, here is how to read the full DevDay slate without getting distracted by the headline act.

What Dots Actually Is

Each dot is an agent instance that runs GPT-6 Astra on a dedicated cloud computer with its own browser. You give it a goal, connect the apps it needs, and set autonomy boundaries — what it can do independently, what needs your approval, and what is off-limits. It persists across sessions, learns from feedback, and integrates with Slack and Teams.

The pitch is compelling on paper. The execution, so far, is not.

During the live DevDay demo, the first dot failed to respond to its prompt. The presenter's recovery — "I guess Dot's having a slow morning" — became the most quoted line of the event. Early testers report that tasks taking Dots 20–30 minutes finish in 5–6 minutes on Claude Sonnet 5.5. The call feature, one of the headline capabilities, did not work during initial rollouts.

Availability tells its own story: Dots launched for Pro and Business Premium subscribers, but explicitly excludes the European Economic Area, Switzerland, and the UK. When your product cannot ship to the jurisdictions with the strongest data protection laws, practitioners notice.

The HN Verdict: 629 Comments, Nobody Asking for Access

View on Hacker News →

The Hacker News thread on Dots is one of the most revealing community reactions to an AI product launch this year. At 750 points and 629 comments, the engagement is high — but the sentiment skew is striking.

The top-voted skeptical comment captures the practitioner confusion: "I genuinely can't work out what Dots actually is. It seems like a dumbed down reskin of Codex/ChatGPT Work but with the power-user features e.g. visibility/mentions removed. As a serious engineer why would I want that?"

Another practitioner identifies the fundamental bottleneck that no agent product has solved: "I have very little need to run Agents overnight, as my throughput is limited by my approval. Each work usually needs revisions, sometimes the bug is just a symptom of the root problem."

This is the approval bottleneck problem that our coverage of agent billing and durable execution has tracked all year. The constraint is not compute or availability — it is human review capacity. An always-on agent that generates work faster than you can review it is a liability, not an asset.

The lone positive thread worth reading: one developer describes running AI "employees" for 6+ months, handling customer support, bug triage, and deployments through dedicated Unix accounts with email integration. But notably, they built this on custom infrastructure — not on a platform product like Dots.

⚠️ Contrarian Corner: What if "dumbed down" is the point?

Power users scoffed at the iPhone for removing the keyboard. They mocked Dropbox as "just rsync with a GUI." The pattern repeats: consumer-first products that expert users dismiss often win precisely because they strip power-user features.

If the always-on agent paradigm normalizes for 1.2 billion ChatGPT users who have never heard of Codex, the developer workflow tools become a trailing indicator. OpenAI's history suggests the consumer bet usually wins — ChatGPT itself was the "dumbed down" version of the API. Worth holding this possibility even while the current product underwhelms.

The Buried Number: 1.2 Billion Weekly Active Users

Here is the figure that most DevDay analysis treated as a footnote: ChatGPT now reaches 1.2 billion weekly active users. That is up from 900 million in February 2026 and 400 million a year before that. OpenAI also disclosed 35 million weekly ChatGPT Work and Codex users and 2.5 million businesses on OpenAI products.

ℹ️ The distribution math that matters: 1.2B WAU is roughly 3x Instagram's user count when Meta launched Threads. It is the largest single-product user base in AI. Every dot, every plugin, every workflow that gets built on this base benefits from distribution that no competitor can replicate through product quality alone.

This is why Dots matters even if the product is mid. OpenAI is not competing on agent capability — Anthropic's Claude arguably wins that fight today, with Sonnet 5.5 delivering 30% faster results at 30% less usage than comparable OpenAI models on agentic coding. OpenAI is competing on distribution. A mediocre agent at 1.2B scale can generate more training data, more plugin integrations, and more network effects than a superior agent at 1/10th the reach.

View original post on X →

For operators, this reframes the competitive question. It is not "which agent is better?" It is "which distribution channel reaches your users?" If your users live in ChatGPT — and statistically, they probably do — Dots becomes relevant regardless of its current quality.

What DevDay Actually Shipped for Builders

While Dots dominated the news cycle, three announcements from DevDay have more immediate implications for agent operators:

Codex Cloud: The Persistent Environment Story

Codex Cloud introduces isolated cloud environments where each coding task gets its own container with the project's repositories, tools, and dependencies. Start a task on your laptop, close it, resume from your phone. Share environments with your team. This is the "always-on" story that actually works for developers — not an anthropomorphized agent, but persistent infrastructure.

For teams already using background agents — and our coverage of fleet orchestration shows this is an accelerating pattern — Codex Cloud's reusable environments with networking and secrets sharing is a concrete upgrade over ephemeral containers.

Ultrafast: 300 Tokens per Second at 6x Price

The new Ultrafast tier generates up to 300 tokens per second in Codex and 6x faster in the API. The pricing is 6x standard rates — so GPT-6 Astra Ultrafast runs $60 input and $300 output per million tokens.

View original post on X →

The operator calculus: if your agent workflow is latency-bound (real-time code review, interactive debugging, customer-facing chat), Ultrafast changes the math on what is worth automating. If your workflow is cost-bound (batch processing, nightly CI runs, fleet operations), it does not help you.

GPT-6.1 Sol: The Price Collapse That Changes Agent Economics

Sol ships at $2/$10 per million tokens — one-fifth of Astra's standard rate — with a 95% cached input discount at $0.10 per million. In planted-bug testing, Sol found 44 bugs for $6.56 versus Astra's 45 for $33. Near-parity performance at 80% cost reduction.

This is the announcement that should change your infrastructure decisions this week. If you are running agents on Astra and paying Astra rates for tasks that Sol handles equally well, you are overspending by 5x. The harness-not-model thesis applies in reverse here: when the model gets cheap enough, the harness economics change.

The Agent Landscape After DevDay

DevDay landed in a week where the agent space is accelerating on multiple fronts simultaneously:

• Anthropic filed for IPO while pulling back API discounts — signaling margin discipline ahead of public markets. If you built on discounted Claude pricing, plan for sustainable rates.
• Manus 2.0 gave agents their own phone number, email, and payment rail — crossing the "agent-as-legal-actor" line that regulators have not addressed.
• The code review bottleneck emerged as the defining problem: AI Engineer conference data showed teams writing 741% more code but shipping only 30% more software. The constraint moved from writing to trusting. This is the problem Dots inherits, not solves.

The pattern across all of these: the agent product is not the bottleneck. The approval, review, and trust infrastructure around agents is. Every new agent capability — whether it is Dots, Claude-as-coworker, or Grok Bot — runs into the same wall. The guardrail stack is where the real work happens.

What Operators Should Actually Do This Week

💡 Try now, wait on, skip — the DevDay operator playbook:

Try now:

• Codex Cloud — persistent environments with team sharing are shipping and useful today. If your team runs background agents, this is a direct upgrade.
• GPT-6.1 Sol — benchmark it against your current Astra or Claude workloads. The 5x price drop at near-parity performance is real. Start with coding tasks where Sol's planted-bug scores match Astra.
• Ultrafast — if latency is your binding constraint (real-time agent UX, interactive coding), test the 300 tok/s tier. If cost is your constraint, skip it.

Wait on:

• Dots — until the EEA exclusion lifts (privacy story is incomplete), memory controls mature (you cannot view or delete individual dot memories), and reliability stabilizes (the live demo crash was not a fluke — early testers report multi-minute hangs).

Skip:

• Dots as your primary agent strategy — building business-critical workflows on a platform that cannot ship to Europe, crashes during demos, and offers no granular memory control is a risk you can avoid. Claude Code, Codex (the developer tool, not Dots), and custom harnesses give you more control for less risk.

The Real Moat Is Not the Agent

OpenAI's DevDay revealed a company playing two games simultaneously. Game one: ship Dots as the consumer-facing agent that turns 1.2 billion ChatGPT users into agent consumers. Game two: ship Codex Cloud, Ultrafast, and Sol as the developer infrastructure that earns the revenue to fund game one.

Practitioners saw through game one immediately — the HN thread, the demo crash, the EEA exclusion all signal a product that is not ready for production workflows. But game two is real, and it is shipping now.

The deeper lesson for the agent ecosystem: distribution, not capability, is becoming the primary competitive axis. Anthropic arguably builds better models. OpenAI indisputably has more users. The question for operators is not which agent is technically superior — it is which ecosystem your users, your compliance requirements, and your cost constraints point toward.

The agents are getting good enough. The question is who gets to distribute them to a billion people first. OpenAI just answered that question, and the answer is not Dots — it is ChatGPT.

---

What is your team testing from DevDay? Join the conversation on AgentConn or reach us on X @ComputeLeapAI.

Originally published at AgentConn