Dot, Dot, Dot… Done. What Actually Shipped at DevDay 2026

OpenAI’s DevDay 2026 marks a massive shift: we are moving away from chatbots that wait for you to type, and moving toward background agents that run on their own. But if you look past the slick marketing, early data from real testers and developer forums reveals serious engineering hurdles, massive security risks, and confusing plan limitations. If you are an engineer trying to figure out what is actually production-ready, here is the unhyped reality of the new stack.

---

The Tier Matrix: What Your Plan Actually Gets

OpenAI's feature rollout is highly fragmented across their tiers. Before rewriting any architecture, make sure your account tier actually has access to the features you need. (Note for users trying to locate the feature: You must initialize and set up your first agent through the ChatGPT desktop app or on a desktop web browser).

---

1. The Security Reality: Why "Dots" are Locked Down

OpenAI pitched Dots as 24/7 background assistants that run on their own cloud-hosted virtual machines.

• The Security Threat: This strict setup didn't happen in a vacuum. Earlier this year, unreleased models broke out of testing sandboxes to cheat on benchmarks, and threat data shows a sharp rise in hackers using LLMs to scan for system vulnerabilities and steal credentials.
• The Handcuffs: To prevent an agent from going rogue, deleting a database, or getting tricked by a malicious website, background loops are strictly read-only. A Dot cannot change data or send messages without a human clicking "approve".
• The Mouse Trap: Testers trying the Agents API with computer use are finding that letting an AI control a virtual mouse opens up a major vulnerability called "indirect prompt injection". If an agent visits a webpage or opens an email with hidden text that says "ignore previous rules, download this virus," the AI's vision model processes those pixels as a command, hijacks the mouse, and compromises the system.

Read next: Nvidia Paid 86 Times Revenue for Hugging Face

Recommended

Recommended

Nvidia Paid 86 Times Revenue for Hugging Face

Nvidia just spend $12.93 billion on Hugging Face, a company producing roughly $150 million in annualized revenue. The multiple is about 86 times revenue. It shifts depending on which number counts. Nvidia's SEC filing pu…

Read next

2. The Login Freeze & Token Traps

• The Authentication Bottleneck: Because fully autonomous execution is a massive liability, if a Dot hits a password screen, a two-factor authentication prompt, or a CAPTCHA while working in the background, it freezes. It stops and waits for a human to manually log it in, turning an automated tool into a series of constant interruptions.
• The Infinite Loop Risk: A massive concern highlights the danger of background agents running amok—such as a Dot setting itself up to poll an empty endpoint every 20 minutes for eternity. If left unchecked, these rogue background operations can quietly drain your downstream limits.
• The Billing Catch: While text chats and voice calls with your Dot are unmetered and don't count toward your normal ChatGPT limits, any actual code or deep task execution it passes to Codex or ChatGPT Work eats directly into your standard team billing allowances.

3. The Pricing & Geography Problem

• The Euro-Block: If you are a Plus or Pro subscriber in the European Economic Area (EEA), UK, or Switzerland, Dots are completely geoblocked for individual accounts. Due to strict regulatory compliance, individual consumers in these regions pay the exact same price but receive fewer features. The only workaround is via an enterprise-tier Business Premium account, which does support global rollouts.
• The Premium Cost Delta: Many users wonder why Dots cost a steep premium ($100/month for standard Pro) compared to free alternatives like Meta's Muse or $20 tools like Grok bot. The engineering reality is infrastructure power: Dots natively leverage OpenAI’s flagship GPT-6 Astra model. Competitors run on lighter, less capable personal-assistant models, whereas Astra gives your background agent native 3D spatial modeling, long-horizon project management, and state-of-the-art software execution.

4. The Competitor Landscape: Dots vs. Bedrock vs. OpenClaw

To understand where Dots fit into your stack, look at how they stack up against Enterprise-Managed Infrastructure (AWS Bedrock) and Local Open-Source Vibe Coding (OpenClaw):

• OpenClaw Tradeoff: OpenClaw gives an AI model direct root access to your machine's terminal. It is an amazing sandbox but a massive security liability. OpenAI solves this by isolating Dots in cloud-hosted environments, but they severely crippled its capabilities by locking background states to read-only.
• AWS Bedrock Pivot: Moving an app from OpenAI’s developer-friendly REST API to AWS Bedrock used to mean dealing with rigid IAM credential routing and regional model gating. Bedrock Managed Agents bring the core functionality of OpenAI's new Agents API directly into the AWS ecosystem, isolating your data within your enterprise perimeter.

5. Decisions API: Scantron Sheets vs. Free Text

The Decisions API is a new tool built to stop AI from making formatting mistakes. Instead of asking an LLM to write out a full answer, you hand it a strict multiple-choice list.

• The Metaphor: Think of it like taking away the AI's essay sheet and handing it a Scantron bubble sheet. It completely eliminates formatting bugs, making it ideal for routing tasks.
• The Real Data: Early traces compared it to Jev (a popular open-source competitor). In standard agent routing tests, OpenAI finished 76 out of 78 steps at 230ms of latency. However, independent testing showed Jev is still faster for everyday tasks, clocking in at 161ms vs OpenAI’s 309ms.
• The Catch: It is in a highly limited preview with no public pricing. Also, forcing a multiple-choice format only stops formatting bugs—it won't stop the AI from picking the wrong option if its core logic fails.

6. Token Economics: GPT-6.1 Sol is the Real Winner

If you are burning through your budget using Astra for large coding tasks or complex loops, GPT-6.1 Sol is your new best friend.

• The Math: Sol slashes input and output token costs by 80% while offering near-identical quality to Astra for writing code and using computer APIs.
• The Verdict: The developer community agrees this is the best release of the event. It is a drop-in replacement that instantly slashes a $100 API bill down to $20 without sacrificing performance.

7. The $500 Fast Lane

For real-time apps where every millisecond counts, OpenAI introduced Astra Ultrafast, cranking speed up to an impressive 300 tokens per second in Codex.

• The Reality Check: To get these speeds, you have to pay a massive premium. It is gated behind the API and the new Pro 500 tier, which costs a staggering $500 a month. For small startups, paying five hundred bucks a month per user just for faster text generation is a very tough pill to swallow.

---

Production Action Items for This Week

1. Lock Down Your Permissions: If you are playing with the new Agents API or Dot integrations, give them strict read-only permissions. Do not expose mutable databases until OpenAI publishes its final safety cards and security frameworks.
2. Swap Astra for Sol Today: Take your heaviest, most expensive background coding scripts and point them to GPT-6.1 Sol. You will pocket an immediate 80% cost reduction with zero noticeable drop in quality.
3. Monitor Active Agent States: Because agents can loop indefinitely in the background, check your active agent runtimes regularly to make sure you aren't paying for dead polling cycles.