OpenAI Dots Explained: Features, Pricing, Examples and Benchmarks

OpenAI’s new personal agents can keep working between conversations. Here is what they can do, how access and costs work, and what the launch evaluations establish.

Published September 29, 2026. Updated October 1, 2026 at 6:56 a.m. PDT. Research based on linked documentation and published evaluations. The example briefs and cost scenarios below are illustrative; Kingy.ai has not run a hands-on dots benchmark.

OpenAI launched dots at DevDay on September 29, 2026. A dot is an individual personal agent inside ChatGPT, powered by GPT-6 Astra. It has its own cloud computer, uses connected tools, retains useful context, and keeps work moving between conversations. OpenAI’s launch announcement

The useful promise is continuity. You can hand over a responsibility, return later with a change, and review the work that accumulated in between. That could mean tracking an API migration, maintaining a research brief, or preparing material for a meeting. The practical questions are whether it finishes your particular work accurately, respects its scope, and saves enough review time to justify the cost.

Some answers are already concrete. Others remain open: permanent post-launch usage terms, measured everyday task costs, and independent end-to-end productivity results. This guide separates those categories so you can make a buying decision without treating a demo as a guarantee.

What is a dot, and how does it work?

The documented system combines Astra, a cloud computer and browser, connected apps, remembered context, and background work. It can research, analyze data, prepare documents, and build software while your own device is off. You can keep talking to it while it works. Meet dots

Think about a software migration. “Find every use of this retiring API” is one task. “Keep the migration moving until all services are off the API” is a responsibility: identify dependencies, coordinate fixes, verify progress, and return with decisions. A useful agent needs to carry the earlier findings forward as new information arrives.

1. Define the jobGoal, sources, boundaries

2. Gather contextFiles, apps, remembered decisions

3. Move work forwardResearch, create, delegate, check

4. Review and steerResults, decisions, corrections

A dot can divide work among parallel background agents, create visible cloud threads, and work through local Work or Codex tasks when a computer is connected. Its context includes relevant ChatGPT memory and its own notes; those notes are not a complete transcript. Proactive research uses restricted read-only tools, with follow-up actions subject to permissions. Tasks and memory

That distinction matters in a business workflow. Finding a stale claim in a proposal is one operation. Editing the shared document or emailing the customer is another. Write the expected deliverable and the allowed actions into the request, especially when a job runs for days.

Availability and setup

At launch, eligible adult Pro users have access outside the European Economic Area, Switzerland, and the UK. Business Premium users can access dots across supported ChatGPT regions. Enterprise access is a beta that administrators must enable; it starts disabled. The rollout may take several days. OpenAI Help Center, Access requirements

The Enterprise beta also covers Edu and Healthcare. Launch announcement

Create your dot in the desktop app or a desktop browser. Name it, review the proposed connections, and give it an initial responsibility. Continue in the mobile app when its supporting update is available; mobile web is unsupported. You can start with a narrow set of connections. Getting started

For a first job, choose something you already know how to check. A weekly change brief is easier to evaluate than a broad instruction to “handle my business.” Supply the source list, the deliverable, the deadline, and the decisions that need you. Give it feedback on the first result before adding more responsibilities.

Product specifications and practical limits

The first four rows summarize Meet dots. The connection details come from Computers and apps. Unspecified rows are gaps in the evidence, rather than claims that no limit exists.

You can inspect the cloud computer and take over its mouse and keyboard. Website sign-in uses a private form or takeover flow. A personal browser login does not sign the dot in, and some websites block cloud browsers or require verification. Local skills require a connected computer. Messaging, app access, and computer access are separate connections. Computers and apps

OpenAI identifies Linux and Chrome as the maintained cloud environment. Sandboxing limits code and tool access; the systems enforcing required checks sit outside the environments a dot can change. That describes the workspace architecture, without establishing CPU, RAM or storage allocations. OpenAI’s security explanation

Astra’s API specifications are useful background, but an API context window does not establish how much information your dot retains or retrieves across months of work.

These are Astra model specifications. Dots’ voice interface can combine additional capabilities; the base model’s modalities do not describe the whole product.

Messaging, voice and workplace connections

You reach the same dot through its available contact methods. Channels retain their own messages while relevant context can carry across them. You can start a voice call and type during it; assigned work can continue after the call ends. Adding a dot to Slack alone does not establish a monitoring schedule. Specify what to watch and confirm the recurring task. Messaging documentation

Treat access labels carefully. The enterprise administration guide describes Teams as an invite-only alpha. It also says only the owner can direct a dot through supported Slack interactions; another participant’s message does not itself start work. Dots administration

Texting, where enabled, is a limited beta for US Pro users, unavailable in Business and Enterprise. Message and data rates may apply. A dot cannot initiate calls or have its own standalone email address at launch, although you can connect a personal email account. OpenAI Help Center

OpenAI advertises connections to more than 4,000 apps through its plugin ecosystem. That is ecosystem reach, not a promise that every account has every integration or action enabled. Launch announcement

Published examples and early hands-on evidence

OpenAI’s product page illustrates several ongoing jobs:

These are OpenAI’s product examples, not Kingy.ai test results. The page does not supply a matched task dataset, completion-time distribution, or invoice-backed cost for each workflow.

The strongest demonstration for your own use case would include the original input, the resulting files or changes, a check against the input, and every correction a person had to make. A polished final slide by itself leaves most of that invisible.

An early hands-on outreach report

On September 30, Promptslove’s Ramanpal Singh published a first-person dots report with screenshots about promoting his own tools. One screenshot shows his dot reporting 13 outreach/submission actions with none of three targeted live links verified; he later reported one live directory listing. Those are activity reports from one campaign, without a controlled success-rate denominator or measured task cost.

Kingy.ai independently checked the public Full Stack Journalism tool directory. It contains Promptslove’s Perplexity Prompt Generator, links to its live destination, and identifies the entry as submitted by its creator for editorial review. The link has a nofollow attribute. This verifies the listing’s presence; we could not independently establish who created it, validate the other outreach actions, or confirm editorial approval or SEO gains. Use the report as a workflow example with a checkable artifact. It does not establish a dots productivity benchmark.

A Japanese setup and research test

Touch AI’s Daisuke Nakajima reports a September 30 desktop-browser test in Japan on Pro 200, with Gmail and GitHub already connected and default low-risk tool permissions. Its screenshots show a dot announcing three email-derived suggestions and returning a four-column launch-summary table. We inspected those screenshots; mailbox details are blurred and browser navigation is not shown. The reported roughly one-minute suggestion interval and sub-two-minute summary are single-run timings without raw traces or metered costs.

The approval check asked the dot how it would handle sending and calendar edits. No email-send action was tested. A model explaining its rules does not verify that a later action will follow them.

A company-paperwork trial

Platformer’s Casey Newton reports using his dot, Kicker, for a few hours while writing his September 29 column (Pacific time). He says it used a lease, city websites and budget documents to answer all but two questions on an office-insurance form. It also drafted a meeting agenda from a transcript and, with his permission, emailed his speaking agent about a scheduling conflict.

Newton estimates that work he would have spent two hours doing required about 15 minutes of his effort. This is his retrospective estimate from limited use. The report does not publish the underlying private documents, complete action traces, a matched manual trial or a metered task charge. It adds a concrete paperwork use case; accuracy and repeatable time savings still need testing.

A first-day email and voice report

In a September 30 first-person report for The Neuron, Corey Noles says setup took roughly five minutes and he continued by voice the next morning. He connected Google Calendar, Docs and Drive, then let his dot interview him about priorities and preferences.

Noles also reports asking it to handle an email: it drafted and sent a reply before he reviewed it. He found the reply acceptable, then instructed it to show outgoing messages first. The report omits the exact request, prior permissions, custom-rule settings and action trace. It cannot establish whether sending exceeded authorization or how often this occurs. His usage-policy instructions were easy to set, but their sustained effect remained untested. This is an early workflow account, without measured task charges or a reproducible benchmark.

Five practical task briefs you can adapt

The following prompts are original, untested examples. They specify what a useful result should look like. Adjust them to the connections and permissions your account supports.

1. Maintain a vendor-change brief

Every weekday at 8:30 a.m. America/Vancouver, review the official changelogs and pricing pages in the attached vendor list through October 9. Keep a dated change log with direct source links. Flag changes affecting our current plan, API behavior, or launch schedule. Put routine updates in this conversation; notify me immediately only for a breaking change or a decision due within two business days. Ask before contacting a vendor or changing an account. Confirm the saved schedule and end date.

The deliverable is a dated record of changes and their effect on your configuration. Check whether each alert identifies an actual change. Include one deliberately unchanged page in the source list to see whether the agent invents news.

2. Turn a bug report into a reviewable fix

Investigate the mobile checkout failure described in this issue. Use the specified repository and test environment. Reproduce the failure, prepare the smallest useful fix, and create a draft pull request with reproduction steps, test evidence and a short explanation. Keep desktop behavior intact. Ask before changing dependencies, committing money, deploying, or merging. Tell me if you cannot reproduce it.

A passing result needs a reproduction or a clear account of why reproduction failed. Run the relevant checks yourself and inspect the changed behavior. “Tests passed” should come with enough evidence to identify which tests ran.

3. Prepare for a customer meeting

Prepare a brief for my meeting with Acme on Thursday. Use the supplied notes, the connected calendar and the specified customer folder. Include the current goal, unresolved questions, commitments we have already made, and three decisions the meeting should settle. Link the source for each commitment. Draft an agenda for my review. Ask before sending messages or moving calendar events.

Check dates, owners, and promised deliverables against the originals. A plausible but invented commitment can cause more work than the preparation saves. A useful brief makes missing or conflicting information visible.

4. Convert an interview into an editorial package

Use this transcript to draft show notes, five clip candidates with timestamps, and three social posts in the attached house style. Distinguish direct quotations from paraphrases. Link each proposed factual claim to the transcript or a supplied source. Save the package for review and ask before publishing. Apply my edits consistently across the package.

For a test, pick a transcript you have already edited. Verify every quotation and timestamp. Count how many proposed clips survive review, rather than counting how many the agent produces.

5. Reconcile a recurring reporting spreadsheet

Update the weekly report from the two supplied exports. Preserve the formulas and column definitions in the template. Reconcile row counts and totals, identify duplicates, and list records that could not be matched. Save a new dated version plus a change log. Ask before overwriting the original or distributing the file. Stop and explain any mismatch you cannot resolve.

Check the arithmetic independently. Include a duplicate record and a missing identifier in the test data. The useful behavior is to flag both and explain their effect on the totals.

How much do dots cost?

There are three separate questions: the subscription needed for access, usage charged when work runs, and the value of the useful tasks that survive review.

Pro prices come from the updated Pro tiers FAQ. Business and Enterprise prices come from Business pricing. These are subscription prices, before applicable tax. Dots’ access documentation names all three Pro tiers for adults and Business Premium; Free, Go, Plus and Business Standard are not listed as eligible launch plans.

Your first dot is included in Pro or Business Premium at no extra cost. OpenAI plans additional dots and options to increase speed or monthly work; prices and timing are unannounced. Launch announcement

New Business workspaces require at least two paid seats, which can mix Standard and Premium. Standard costs $25 monthly or $20 per month billed annually. One Premium plus one Standard therefore totals $150/month with monthly billing, or $1,440/year with annual billing. This is our arithmetic for that seat mix, not a standalone dots price. Business signup requirements, Business pricing

Conversations with your dot do not count toward ChatGPT limits. Tasks it starts or manages in Work or Codex count toward those products’ limits normally. Deeper work has an included allowance with extended launch-month limits. Dots usage and access

The Help Center has replaced its earlier blanket launch-month waiver with that allowance description, resolving the earlier wording discrepancy. A numeric permanent deeper-work allowance remains unspecified. Launch FAQ

No fixed dollars-per-dot-task tariff or permanent post-promotion task allowance was established in the checked sources. Buying a subscription also does not establish how many useful tasks you will finish. Failed runs, corrections, connected-service fees and human review belong in that calculation.

A transparent subscription cost calculation

Suppose you allocate a whole subscription to this workflow. Divide its monthly price by the number of completed tasks you accept:

These are arithmetic scenarios, not usage entitlements or measured dots results. They exclude additional usage, taxes, other services and review time. If you already pay for ChatGPT, the incremental subscription cost of trying an included feature may be zero; allocating the whole plan price to dots answers a different budgeting question.

Add human time for a more useful estimate. Assume a $200 subscription, 100 accepted tasks, five minutes of review per accepted task, and labor valued at $60/hour. Review costs $5 per task. Subscription allocation adds $2, giving $7 per accepted task before retries, setup and other charges. Those are declared assumptions, not a claim about dots speed.

If the same job takes you 20 minutes manually, its assumed labor cost is $20. That suggests a possible $13 saving. Change the review time to 18 minutes and the saving disappears. This is why a task’s correction burden matters as much as its apparent completion rate.

Astra API costs provide background, not a dots bill

Developers using Astra directly pay for tokens. Standard short-context rates are $10 per million uncached input tokens, $1 per million cached reads, $12.50 per million cache writes and $50 per million output tokens. API pricing

Above 272,000 input tokens, the entire request uses doubled input/cache rates and 1.5 times the output rate. Astra model page

Billable output includes reasoning tokens, which may never appear in the final answer. Reasoning documentation

The token counts here are invented inputs for transparent arithmetic. They are not measurements of a research brief, bug fix or dots task. The examples exclude tools, cache writes, multiple calls, retries and external services. An agent can make many model calls while completing one human request, so a single-request price cannot establish an end-to-end task price.

Benchmarks and evaluations

Read Astra capability benchmarks, dots-specific safety tests and public workflow comparisons separately. They measure different things.

Astra’s capability benchmarks

OpenAI reports maximum scores across reasoning efforts in research or API environments. Prompts and tools can differ from production ChatGPT. These are base-model results, not dots completion rates. Astra launch benchmarks

A benchmark score can help identify an underlying model’s strengths. It cannot tell you whether a connected calendar is configured correctly, whether a website will allow the cloud browser through, or whether a generated presentation matches your reporting definitions. Those dependencies are part of the product experience.

Dots-specific safety evaluations

OpenAI added a dots appendix to the Astra system card on September 29. Selected results:

These are OpenAI-run evaluations under specific conditions. Some tests isolate behavior without the full production safeguards. Zero observed successes does not establish immunity, and the boundary failures deserve attention. These numbers measure safety behavior, not everyday completion or uptime. Astra system card, dots appendix

For a buyer, the actionable implication is to make authorization changes explicit. When a project switches from preparing a draft to making a live change, state the new boundary. Revisit persistent responsibilities when their owners, data sources or intended audience change.

A competitor’s five-task comparison

Retriever AI’s comparison, tested September 29, 2026 (Pacific) on gym-0.3.0, reports dots passing 22/24 objective checks and rtrvr 24/24 across fictional flight-credit, job, creator, invoice and inbox tasks. The publisher sells rtrvr. Its selected runs used GPT-6 Luna and GLM 5.3 Flash; dots’ model and reasoning effort were not exposed. Conversation reset and memory settings were not recorded.

The invoice difference needs context. Dots’ saved report contains the correct $330 total and identifies all three exceptions, but fails two strict field-format checks. Rtrvr also initially scored 3/5, then used feedback to reach 5/5. We inspected the linked receipts and public evaluator; the receipts are editable, and we have not independently replicated the runs.

The page’s approximately $20 dots figure allocates reported 10% usage of a $200 subscription; rtrvr’s $0.0709 counts model tokens. These incompatible bases do not establish a 300× task-price advantage. Reported timings remain unaudited. This small public test provides checkable artifacts, with limited evidence for real-world reliability.

What remains unmeasured

The sources checked for this article do not establish an independent dots benchmark for accepted work per dollar, multi-day completion reliability, average correction time, or p95 task latency. Kingy.ai has not performed those tests. Launch examples and model scores should not fill those gaps by implication.

How to evaluate a dot on your own work

Start with ten historical tasks whose answers or acceptable outputs you already know. Give the dot the original inputs, without your completed solution. Use the same scope and grading rules for a manual baseline or another tool. Keep time spent setting up connections separate from time spent reviewing each result.

Score each result as accepted, accepted after correction, or rejected. Record elapsed time, review minutes, rework minutes, observed usage debit, additional fees and any action outside scope. Preserve the output that was graded. A vague overall satisfaction score makes it hard to diagnose failures.

Calculate cost per accepted task as allocated subscription cost plus metered extras, connected-service costs and human review/rework, divided by accepted tasks. Record rejected-task costs too. Compare that figure with the manual baseline. Run a few tasks again on another day to check consistency before relying on recurring work.

For long jobs, include a change halfway through: revoke a permission, revise a deadline, or correct an earlier assumption. Grade the subsequent actions. A persistent agent needs to respond to changed conditions as reliably as it handles the initial prompt.

Permissions, memory and stopping work

Automatic review checks consequential actions against instructions, permissions, custom rules and safeguards. Drafting does not authorize sending. Rules can require approval or hand a step to you, but do not grant app access or override safeguards. Open Activity to inspect delegated results and Scheduled to inspect recurring work. Controls documentation

Stopping has separate layers. Pause stops the main task; stop delegated tasks through Activity and cancel recurring runs through Scheduled. Completed actions are not undone. OpenAI says proactive research and private notes are not directly used for training; eligible conversation and task information follows ChatGPT data settings. Controls documentation

Disconnecting an app does not erase information already obtained. Resetting deletes the dot, its conversations, saved memories and scheduled tasks. OpenAI Help Center

Reset has limits: separately stored files, Codex threads and ChatGPT conversations created by the dot remain, as do other ChatGPT memories. OpenAI says you cannot currently view, directly edit or individually delete the dot’s own memories. Turning off ChatGPT Memory stops further memory sharing; it does not erase information already received. Business, Enterprise and Edu content is excluded from model training by default. On personal plans, the “Improve the model for everyone” setting governs eligible conversations and work. Privacy FAQ

OpenAI says dot context excludes credentials, images and screenshots; files in ChatGPT Library, local storage or connected services follow that location’s retention rules. Turning off model improvement does not rule out limited human review, including safety cases. Advance permission for messages covers only the specified recipient, content and timing or trigger; approving one message does not authorize continuing outreach. Privacy FAQ

Proactive research cannot directly contact people, change connected-app content or control a browser or desktop; OpenAI says code enforces these restrictions. Follow-up actions require the usual checks. Mandatory confirmations include permanent deletion, software from unrecognized sources and new security-sensitive access. Password changes and transfers between financial accounts require you to take over. Custom Rules cannot remove these safeguards. OpenAI’s security explanation

Administrators have another important check: enterprise model defaults and controls do not apply to dots. The administration guide also restricts local access when the applicable residency policy requires it. Verify the actual workspace configuration before assuming a familiar policy covers the new feature. Dots administration

Specialist dots and what is still coming

OpenAI describes specialist dots for organizational responsibilities, including procurement, invoice processing, customer support and commercial contracting. These are focused enterprise pilots. Microsoft Agent 365 integration is described as a coming capability. Launch announcement

An enterprise pilot does not establish general self-service availability. Ask for the supported workflow, accountable owner, data boundaries, pricing and rollout status for the specific specialist you want.

When dots are worth trying

A dot is a sensible candidate when work recurs, context accumulates, and the output can be checked. Research monitoring, document preparation and a migration with reviewable changes fit that pattern. The likely value comes from reducing repeated coordination as well as producing the artifact.

For a short, isolated question, an ordinary conversation may be enough. For a narrowly specified coding task, directing Codex yourself may give you the control you need. Use a dot when you want an ongoing responsibility coordinated across tasks and tools. These are workflow recommendations, not comparative performance results.

Start with one responsibility and an acceptance standard. Track the useful results, the corrections, and the full cost through the launch period. Recheck OpenAI’s usage terms before budgeting for the following month.

Update log

ShowHide

• September 29, 2026, 2:36 p.m. PDT: Added included-dot and expansion details; clarified Edu/Healthcare beta access from the September 29 launch announcement.
• September 29, 2026, 8:37 p.m. PDT: Updated usage guidance after the Help Center removed its blanket waiver; retained normal delegated-task limits from Learn.
• September 30, 2026, 8:46 a.m. PDT: Added cloud workspace and enforced research restrictions from OpenAI; clarified memory controls, reset limits and training settings using the Privacy FAQ.
• September 30, 2026, 12:50 p.m. PDT: Added context-retention details, limited human-review disclosure and the boundaries of advance message approval from the revised Privacy FAQ.
• September 30, 2026, 2:50 p.m. PDT: Added an attributed Promptslove hands-on outreach report and our check of its public directory artifact, with limits on what the evidence establishes.
• September 30, 2026, 4:51 p.m. PDT: Added Retriever AI’s five-task comparison, explaining invoice grading, evidence limits and incompatible cost measures.
• September 30, 2026, 8:50 p.m. PDT: Added Touch AI’s Japanese hands-on report, with inspected screenshots and explicit limits on timing and approval evidence.
• October 1, 2026, 12:54 a.m. PDT: Added Platformer’s company-paperwork trial, separating reported outcomes and estimated time savings from measured benchmarks.
• October 1, 2026, 6:56 a.m. PDT: Added The Neuron’s first-day report, including setup, voice and an email sent before review; identified the missing authorization evidence and untested long-term preferences.