What a Dot Cannot Do Yet: The Honest Limits List

Launch coverage for a new AI product almost always reads as a capability list — what the thing can do, illustrated with the best-case demo. The more useful document, and the one vendors never publish themselves, is the limits list: what's genuinely off-limits by design, what's simply not built yet, and what's been demoed but not verified at scale. Nine days after OpenAI introduced Dots at DevDay, here's the honest version of that list, drawn only from what's been confirmed across the coverage of the September 29 launch.

Explicitly restricted by design

Some limits are deliberate and disclosed. Proactive research — a dot's ability to work in the background on projects even when it's not actively tasked — is explicitly confined to read-only tools that cannot send messages or modify content. That's a genuine safety boundary, not a placeholder: OpenAI built proactive research to be incapable of taking consequential action on its own, only surfacing findings for a human to act on.

Sensitive operations, like password changes, remain reserved for humans regardless of what permissions a user grants a dot elsewhere. That's a hard line OpenAI has drawn rather than left to Custom Rules configuration, and it's worth noting as one of the few unambiguous "a dot cannot do this, full stop" facts in the entire launch.

Not yet shipped

Several capabilities referenced in launch coverage are explicitly future-tense rather than live. Text message and phone/audio access are "coming soon," not available at launch — meaning a dot's ability to reach a user outside of ChatGPT, Slack, or Teams doesn't exist yet, a real gap for the "keeps working and returns when things change" pitch. Slack, Teams, and Soon Your Phone covers that gap directly. Team configurations for running multiple dots together are also described as "coming later" rather than available now, despite "running multiple dots as a team" being referenced in the same announcement — the feature and its general availability date are two different things. And collaborative slides and spreadsheets inside ChatGPT Space are listed as "coming" rather than shipped, alongside the Pages feature that is live.

Undisclosed rather than restricted

This is the largest category, and it's worth being precise about the distinction between "undisclosed" and "doesn't exist." OpenAI has not published: the numerical usage allowance for intensive dot work beyond the first free dot; the eventual price of additional dots; pricing for higher speed or workload capacity tiers; or commercial pricing for Specialist Dots specifically. None of that means these things don't exist or won't be reasonable — it means anyone evaluating Dots for real budget planning right now is working from a genuine information gap, not a minor omission. Pricing the Unpriced covers what can and can't be reasonably inferred about where that pricing might land.

The difference between "a dot can't do this" and "OpenAI hasn't said what this costs yet" matters enormously for anyone deciding whether to build a workflow around one — and the launch coverage mostly doesn't bother to distinguish them.

Demoed but not independently verified

The use cases OpenAI showcased — a developer's dot watching feedback and filing PRs, a sales dot comparing prospects against documentation, a content dot propagating transcript edits across social posts — are real product capabilities, not vaporware; the underlying infrastructure (GPT-6 Astra, isolated cloud computers, the plugin ecosystem, Auto-Review) genuinely exists and genuinely works in the demos shown. What hasn't been established by any source reviewed for this series is how these hold up against messy real-world inputs at scale, over weeks of continuous use, rather than a curated DevDay walkthrough. The Developer's Dot, The Sales Dot, and The Content Dot each cover their respective use case's specific unverified claims in more depth.

The comparison worth keeping in view

None of this is a knock specific to OpenAI — Grok Bot has no confirmed formal safety documentation in the sources reviewed for this series at all, and Meta's Muse hasn't yet shipped its promised Confidential VM with end-to-end encryption, targeted for year-end 2026 but not live as of this week. Every always-on agent launched in this eight-week collision is, to some degree, ahead of its own documentation. That's a normal feature of a genuinely new product category moving fast, not evidence of bad faith — but it does mean the responsible way to evaluate any of these three products right now is to separate what's confirmed, restricted, unshipped, and undisclosed, rather than taking a launch announcement's confident tone as a complete account of the product's actual current state.

Why this list is worth revisiting in three months

A limits list like this one has a short shelf life by design — that's the point of writing it down rather than trusting memory of a launch demo. Team configurations, phone access, and Specialist Dots pricing are all things OpenAI has said are coming, not things it's ruled out, and a fair re-check in a few months should specifically test whether each item moved from this list into the confirmed-shipped column, or whether it quietly slipped further out. That kind of tracking is a more useful long-term habit for evaluating agentic AI products generally than treating any single launch announcement as the final word — these products are genuinely moving fast, and the gap between "announced" and "reliable in production" is precisely where most of the real risk, and most of the real opportunity, actually lives. This series has tried to keep that distinction explicit article by article; Proactive Research picks the thread back up with the one capability in this launch that comes closest to living up to its billing.

Part 20 of 30 in the ChatGPT Dots series. Previously: Specialist Dots: The Early Pilots Nobody's Talking About Yet. Next: Proactive Research: The Feature That Makes Dots Feel Like Employees. Browse more builder insights or explore AI skills at aiskill.market.