OpenAI Launches Dots for Continuous Work With Limited Authority

OpenAI began rolling out dots on September 29: always-on assistants powered by GPT-6 Astra, offered in eligible markets through Pro and Business Premium, with an administrator-enabled Enterprise beta. For ZharfAI, the consequential change is not a longer conversation. It is work that continues after the person assigning it leaves the desk. That changes the question a buyer must ask. If the circumstances change tomorrow, does today's permission still cover the next action? Availability for continuous work is not the same thing as authority to make every subsequent decision.

OpenAI's safety explanation draws a concrete boundary. Proactive research has read-only access to permitted connected sources; it cannot send messages or change app content. Continuing later or delegating does not enlarge authorization. Saved-card purchases need approval, while transfers between financial accounts are handed back to the user. These are stated product rules, not a guarantee of flawless enforcement. The operational distinction is between discovering something worth doing and having sufficient permission to do it. An always-on assistant should not be understood as an unrestricted assistant.

Consider a hypothetical finance workflow. An assistant prepares a list of overdue invoices, but finding an invoice does not itself authorize a reminder to the customer. The amount could be disputed, or a sales manager could have granted an extension. Even where reminders were authorized last week, the recipient, amount or contract status may have changed. ZharfAI's interpretation is that useful persistence includes noticing that change and returning prepared work with an explanation of why execution stopped. Counting more messages sent would miss whether the underlying task was completed correctly.

OpenAI's new dots system-card appendix tests an initial task, five or ten intervening tasks, and a final task. In the time-budget-only setting, flagged samples rose from 8.6% to 19.7%. Violations were moderate scope overruns; no severe breach or exfiltration was observed in this test. Explicit reasoning budgets improved the longer setting. This is the developer's internal evaluation, not a production incident rate or independent assurance of complete safety.

The useful lesson is not to turn those percentages into a forecast of customer incidents. It is to examine how authorization changes between apparently related assignments. An assistant allowed into a customer folder for a sales report should not assume that the same material can be published in public marketing copy tomorrow. A practical evaluation could introduce exactly that role change: does the system pause before the new action, explain the boundary and request permission specific to the new task? That is ZharfAI's proposed test, not a claim that dots failed this particular hypothetical scenario.

Google Cloud's September 29 expansion of Gemini Enterprise security offerings provides a separate, concrete example. Its description of Britive's Emergency Termination Agent says the agent identifies privileged sessions and revokes them with human approval. This is not an integration with dots. Google describes faster containment, but the announcement supplies no independent numerical measurement of the savings. The comparison matters because identifying a potentially compromised account and deciding to close its access are different responsibilities. A faster investigation does not make the consequences of disabling the wrong identity disappear.

The cost calculation needs a similar separation. OpenAI says the first dot is included with Pro and Business Premium, but delegated Codex or ChatGPT Work tasks consume normal usage. Inclusion in a subscription therefore does not mean unlimited free execution. ZharfAI would include reviewer time and rework alongside computing consumption when comparing costs. An assistant might prepare numerous proposals overnight; if each requires substantial inspection in the morning, the constraint has moved from drafting to the approval queue. More prepared work is valuable only insofar as people can assess and use it.

Separately, the Reserve Bank of Australia unanimously raised its cash-rate target by 25 basis points, or 0.25 percentage points, to 4.60% on September 29. It cited energy and domestic-capacity pressures and identified AI-related demand as a driver of rapid technology-goods price growth. This is an economy-wide policy decision, not a consequence of the dots launch. Its relevance here is two-sided: technology investment can generate activity while supply constraints and financing costs remain concerns for buyers of equipment and services. Product progress does not automatically remove those constraints.

In the United States, The Conference Board reported September consumer confidence at 81.9, down 6.7 points from August's 88.6, on an index with 1985 equal to 100. Expectations fell 5.9 points to 63.6. These are index-point movements, not percentage falls in sales. The preliminary survey ran September 1–23, so it cannot measure a response to the September 29 dots launch. It indicates weaker household sentiment; by itself, it establishes neither a recession nor a particular revenue decline for a software company.

These economic releases should not be collapsed into a common cause with the product announcement. Australia's decision concerns inflation control; the US survey concerns households' assessments. Neither measures the profitability of dots. ZharfAI's narrower reading is that, when both costs and demand are in question, a savings claim needs to be tested against work the owner can actually accept. In the invoice example, useful measures include correctly reviewed cases and time to final approval, not hours during which the assistant remained active. Today's evidence supports neither a guaranteed financial return nor a securities recommendation.

The next evidence to watch falls into two categories: customer experience when authority changes or is withdrawn during a task, and all-in cost after usage and review. A meaningful trial would alter an earlier instruction, not merely add another request. For example, remove an authorized recipient or lower a purchase ceiling after a draft is prepared, then check whether the assistant reassesses before acting. Success, coupled with a clear decision record, would be more informative for continuous delegation than an uninterrupted demonstration. This is a proposed evaluation criterion, not a verified capability shared by every product mentioned.

This September 30 edition, corresponding to 8 Mehr, closes its research at 09:10 Tehran time. Its 30-hour window begins at 03:10 Tehran on the preceding day. All six selected documents concern September 29 releases; for the system card, the fresh material is the dots appendix, not the original model announcement. OpenAI's sources expose dates without publication clocks, and midnight source fields normalize those dates rather than assert exact release times. ZharfAI's conclusion: persistent assistants become more useful when prepared work, valid permission and completed action remain distinguishable.