OpenAI's Dots give Pro users an agent with its own computer
DevDay 2026 ran an hour and shipped more than 20 announcements. The one that matters is the one that stops waiting for you to type.
OpenAI launched Dots, always-on agents that run on their own cloud computer and keep working when you walk away — powered by GPT-6 Astra and reachable from ChatGPT, Slack and Microsoft Teams. Each dot gets a name, a blob-shaped avatar, its own browser and read-only access to more than 4,000 connected apps for the background work OpenAI calls "proactive research," with an activity view so you can inspect what it did while you weren't looking. Sam Altman's pitch on stage was that Astra is "our most aligned model" — the bet being that alignment, not features, is what earns you the right to hand over an inbox rather than a chat window.
The Meta comparison is unavoidable: same blob avatar, same always-on pitch, same competitor framing in every write-up. But Muse's weekend is the cautionary tale, not the template — Meta's Muse gave a stranger a YouTuber's home address, and Dots launch straight into that trust gap. OpenAI's answer is auto-review before actions that touch your accounts, custom rules that allow, gate or block specific moves, read-only tools during background work, and a cloud computer kept separate from yours unless you explicitly connect your laptop. Sensitive tasks like changing a password always stay with the human. How much that helps depends on whether a permission model can tell the difference between "share my address with a buyer" and "share my address" — the question Monday's incident raised and nobody has answered yet.
GPT-6.1 Sol landed as the cheap route to near-Astra intelligence: $2 per million input tokens, $0.10 cached, $10 output — one-fifth of Astra's $10 and $50, and it is not in ChatGPT Chat yet, only ChatGPT Work, Codex and the API. The cached-input price is the number to keep: 95% below standard input and half of GPT-6 Sol's cache rate, which is what makes an always-on agent that re-reads the same context every turn affordable. OpenAI's own evaluations put Sol level with Astra on DeepSWE 1.1 at roughly a fifth of the cost, ahead of Opus 5.5 on the professional-document and workflow benchmarks at under half the cost per task, and within 2.1 points of Astra on OSWorld 2.0 computer use at about a seventh of the cost. Factual errors on deliberately error-inducing prompts fall from 11.4% to 7.7% at low reasoning effort — a 32% reduction, though those prompts are not representative of normal use.
OpenAI also added a Pro 500 tier at $500 a month, and cut the $200 plan's allowance while it was at it. Pro 500 carries the company's highest usage allowance at 25 times ChatGPT Plus and the only route to Ultrafast, the new speed tier at up to eight times standard token generation — about 300 tokens per second in Codex, at six times the standard price. That is $6,000 a year, roughly double the top consumer tier at Anthropic, Cursor or Google, all of which sit between $200 and $250. Existing Pro 200 subscribers saw the flip side: Codex and ChatGPT Work usage dropped from 20 times the Plus allowance to 10, and GPT-6 Pro messages from 200 to 100 a week, softened by a one-time transition credit. Codex engineering lead Thibault Sottiaux framed it as "[netting] out at half the dollar in API spend compared to the prior plan." Selling speed back to users whose allowance you just trimmed is a pricing-model statement, not a feature.
What to watch: whether Dots reach the EU at all — OpenAI says "eligible markets" and declines to list them.
Is an agent with its own computer and its own identity a tool you'd hand your inbox to, or a liability you'd rather keep on a leash? Tell us in the comments.
