OpenAI launches another milestone with Dots, and full stack AI engineers will feel it first

On September 29, 2026, at DevDay in San Francisco, OpenAI launched Dots. A dot is an AI agent that stays switched on. Each one has its own cloud computer and browser and keeps working on a task list after you close the chat. ChatGPT Pro ($200 a month) and Business Premium ($125 a month) include one dot at no extra charge. The model behind them, GPT-6 Astra, came out a week earlier.

I freelance on Upwork and Fiverr, so I read the launch posts with one question in mind: which jobs that clients pay freelancers for can a $125 subscription now do alone? This article covers what Dots are, how the model was built and tested, the dot.com story involving Grok Bot, how Dots compare with Meta's Muse, and what the marketplace numbers say for people who sell automation work.

Cover graphic: one red dot linked to six smaller dots, representing a team of always-on OpenAI agents, with the article title.OPENAI DOTS · DEVDAY 2026Another milestone forfull stack AI engineersGPT-6 Astra, Grok Bot, Meta Muse and the freelance marketZubair Hussain ShahSeptember 30, 2026

What OpenAI launched

A dot runs on an isolated cloud computer with its own browser. It stays separate from your laptop unless you choose to connect it. You give it work in ChatGPT on desktop, web or mobile, in Codex, or from Slack and Microsoft Teams, and OpenAI lists text messaging as coming soon. One dot can keep several projects going at once, and it learns your preferences from the feedback you give it.

OpenAI says dots can use more than 4,000 apps through plugins. They also do what OpenAI calls proactive research: background reading with read-only access to your apps, so a dot can come back with something before you ask. Sam Altman described them as "the real deal version of AI" people have imagined from movies.

Terms you will see in the Dots docs

Always-on agent

An agent that keeps running after you leave, until the task is finished or you stop it.

Cloud computer

The isolated machine and browser each dot works in.

Plugins and connectors

The 4,000+ app integrations a dot can call.

Custom Rules

Your own list of actions a dot may take alone, may never take, or must ask about first.

Auto-review

An extra check on actions that affect your accounts. Sensitive steps such as a password change always need your approval, and a monitoring system can pause or stop a dot if it spots a safety problem.

Specialist dots

Enterprise dots with a fixed job and their own credentials. The pilot covers procurement, invoicing, email marketing, support and contracting.

Teams of dots

Several dots coordinating on one goal. OpenAI presents this as where the product is heading.

Conditions at launch

Business Premium users get Dots in every region ChatGPT supports. Pro users in the EEA, Switzerland and the UK are excluded for now. Enterprise, Edu and Healthcare workspaces get a beta once an admin turns it on, and OpenAI is working with Microsoft to plug Dots into Agent 365 security controls.

Diagram: you assign a task, Custom Rules check it, the dot runs GPT-6 Astra on an isolated cloud computer and browser, uses 4,000+ app connectors, asks for approval on sensitive steps and reports back in ChatGPT, Slack or Teams.How a dot handles a taskBased on OpenAI's DevDay 2026 launch detailsYou or your teamassign workCustom Rulesallow, block, askYour dotGPT-6 AstraCloud computerown browser, isolated4,000+ appsvia pluginsSensitive step?asks you firstReports backChatGPT, Slack, Teams, SMS soon

GPT-6 Astra, the model inside every dot

OpenAI released GPT-6 Astra on September 22, 2026, first to a small group of organizations and then to Plus, Pro, Business and Enterprise users. API pricing is $10 per million input tokens and $50 per million output tokens. A fast mode runs quicker at twice the price. At DevDay OpenAI also announced GPT-6.1 Sol, which it says comes close to Astra at one fifth of the standard token price.

How models like Astra are built

OpenAI's own description is one line: Astra "brings together years of research and big bets across pre-training, reinforcement learning, and alignment." It does not publish parameter counts, training data size or compute. Here is what those stages usually involve:

1. Pre-training. The model learns to predict the next piece of text across a huge dataset. General knowledge and language ability come from this stage, and it uses most of the compute.
2. Reinforcement learning. The model practices tasks where success can be checked automatically, such as passing tests, running commands or finishing a job in a sandboxed browser, and is rewarded for correct results. For an agent model this is where it learns to click, type, call tools and recover from its own mistakes.
3. Alignment and safety training. The model is trained and tested to follow instructions, refuse harmful requests and stay inside the limits it is given. OpenAI reports a 2.4% error rate on its internal computer-use safety benchmark, down from 22.0% for GPT-5.6 Sol.

That last stage matters more for Dots than for a chatbot, because an agent with hours of access to your inbox has more chances to do something you did not ask for. OpenAI says Astra reaches the Critical level for cybersecurity in its Preparedness Framework. During testing it found two previously unknown zero-day vulnerabilities, which OpenAI disclosed to the maintainers, and some advanced security tasks are restricted at launch.

Benchmarks and how to read them

These are OpenAI's published numbers against its previous flagship, GPT-5.6 Sol. Where OpenAI or DataCamp listed a Claude Opus 5 score, I added it.

Bar chart of OpenAI-reported scores. OSWorld 2.0 72.6 vs 65.7, Agents' Last Exam 59.3 vs 53.6, ScreenSpot-Pro 92.7 vs 76.9, AutomationBench 41.4 vs 18.1, Terminal-Bench 4.0 57.9 vs 37.3, DeepSWE v1.1 74.1 vs 72.7.GPT-6 Astra vs GPT-5.6 Sol on agent and coding testsGPT-6 AstraGPT-5.6 Sol0%25%50%75%100%OSWorld 2.072.6%65.7%Agents' Last Exam59.3%53.6%ScreenSpot-Pro92.7%76.9%AutomationBench41.4%18.1%Terminal-Bench 4.057.9%37.3%DeepSWE v1.174.1%72.7%Source: OpenAI, GPT-6 Astra launch post, September 22, 2026. Vendor-reported scores.

AutomationBench is the number I would watch. It more than doubled, from 18.1% to 41.4%, and it is the test closest to what automation freelancers sell. It also means Astra still failed most of the tasks in it. The computer-use scores (OSWorld 2.0 and ScreenSpot-Pro) explain why a dot can work inside apps that have no API. Astra does not win everywhere: on Humanity's Last Exam with tools it scored 57.2% against 65.0% for a rival model in DataCamp's comparison. Every figure here is vendor-reported and was run with OpenAI's own settings, so wait for independent runs before you quote them to a client.

The dot.com twist and Grok Bot

Hours after the launch, people noticed that dot.com does not lead to OpenAI. It forwards to the download page for Grok Bot, the agent from Elon Musk's SpaceXAI. (SpaceXAI is the former xAI, which was folded into SpaceX in February and renamed in July.) Registration records reported by Cybernews show dot.com was registered on July 28, 2026, with Musk as registrant and an expiry date in 2035. No official reason has been given, and SpaceXAI did not answer requests for comment. The plural, dots.com, belongs to a former fashion business, so OpenAI owns neither.

Grok Bot launched on August 11, 2026. It is an always-on agent on a persistent cloud computer that signs into your software, inboxes and websites, including services with no API. It can watch you do a workflow, save it and repeat it, and several bots can split up a job. You get it through Cursor Premium Teams ($120 per seat per month), Cursor Ultra ($200 a month) or SuperGrok Heavy ($300 a month). SpaceXAI has not named the model, and early tester Matt Shumer noted that users cannot choose or pin it.

Dots vs Muse vs Grok Bot

Muse went to the top of Apple's App Store as a consumer agent for shopping, bookings, bills and forms. Dots aim at work: Slack alerts, invoices, schedules, customer feedback, code fixes. Grok Bot sits between the two. It is sold to developer teams through Cursor, and it leans on logging into websites the way a person would.

Animated 12-second loop. A red dot moves through six steps: task sent in Slack, rules check, work on the cloud computer, apps used through connectors, approval requested, result reported.One task, start to finishAnimated loop, 12 seconds, made in SVGTask in Slackyou askRules checkallowed?Cloud computerdot worksConnectorsapps usedApprovalif sensitiveDonereport sentYou post in Slack: "check new invoices and flag late ones"Custom Rules decide what the dot can do aloneThe dot opens its own browser on its cloud computerIt reads your accounting and email apps through connectorsPaying or changing an account? It stops and asks youIt posts a summary back in Slack and keeps watchingLoops automatically

Text version of the animation

1. You post a task in Slack. 2. Custom Rules decide what the dot may do alone. 3. The dot opens its browser on its cloud computer. 4. It reads your apps through connectors. 5. If the next step pays money or changes an account, it asks you. 6. It reports back in Slack and keeps watching.

What this means for n8n and automation freelancers

Freelancers who sell small, repeatable automations are the most exposed. The launch examples read like a list of common Upwork gigs: watch a Slack channel for alerts, process invoices, update a schedule, go through customer feedback and ship a fix. A client who used to pay for a custom n8n flow for each of those can now describe the job to a dot.

What the marketplace data shows

• Upwork's In-Demand Skills 2026 report (February 4, 2026) found demand for skills tied to applying AI grew 109% year over year. AI integration grew 178% and AI chatbot development 71%, and n8n experts appear in its AI and automation hiring list.
• Fiverr's Business Trends Index (September 13, 2026) shows n8n AI automation up 125%, AI voice agents up 49% and Claude Code up 938%.
• Fiverr's Q2 2026 results (July 29, 2026) show the other side. Annual active buyers fell 21.9% to 2.7 million, revenue fell 10% to $97.8 million, and full-year revenue guidance dropped to between $356 million and $372 million. Spend per buyer rose 15.6% to $368.
• On Freelancer.com, the n8n jobs page listed 7 open projects on September 30, 2026, with average bids from $115 to $768. The AI agents page listed 11.

Bar chart: Upwork AI integration +178%, Fiverr n8n AI automation +125%, Upwork AI skills overall +109%, Upwork AI chatbot development +71%, Fiverr AI voice agents +49%, Fiverr annual active buyers -21.9%.Demand for AI automation is up. Buyers are down.Year-over-year change reported by the platforms in 2026AI integrationUpwork, Feb 2026+178%n8n AI automationFiverr, Sep 2026+125%All AI skillsUpwork, Feb 2026+109%AI chatbot developmentUpwork, Feb 2026+71%AI voice agentsFiverr, Sep 2026+49%Active buyersFiverr Q2 2026−21.9%Sources: Upwork In-Demand Skills 2026, Fiverr Business Trends Index, Fiverr Q2 2026 results.

Demand for AI automation is rising while one of the biggest marketplaces is losing buyers. The buyers who stay spend more each, so the small one-off gig is shrinking and larger engagements are growing.

How much could a freelancer lose?

Nobody has published a loss figure for Dots yet, since they launched this week. You can size your own exposure with Upwork's price guide for AI agent work: simple chatbot setup $500 to $2,000, workflow automation $2,000 to $5,000, a custom LLM agent $5,000 to $15,000, and multi-agent systems $10,000 to $30,000.

A worked example: you sell six simple workflow automations a year at $2,500 each, $15,000 in total, and they look like the launch demos (alerts, invoices, schedules). Your client can get one dot for $125 a month, or $1,500 a year, bundled with ChatGPT. If half of those clients try a dot first and never come back, you lose three projects, which is $7,500 a year. That is an illustration, not a forecast, so run it with your own numbers.

The work that holds up is the work a dot does badly or is not allowed to do. That includes self-hosted n8n for clients who cannot send data to OpenAI, integrations with old internal systems, cost control on high-volume flows where per-token pricing adds up, projects for clients in the EEA, Switzerland and the UK where Pro users cannot get Dots yet, and setting up and auditing the rules, approvals and credentials a team of dots needs.

Rankings on Upwork, Fiverr and Freelancer.com

None of the three platforms publishes a ranked list of top n8n or AI agent freelancers. What they do show is how talent gets surfaced and where demand is moving.

• Upwork's AI agent developer hiring page shows an average of 4.8 out of 5 from 12,625 client reviews and typical rates of $30 to $150 an hour. Placement in search leans on your Job Success Score and badges such as Top Rated, which come from finished contracts and client feedback.
• On Fiverr, gig placement is influenced by seller level and recent order performance. The Business Trends Index is the best public signal of which search terms are growing, and n8n AI automation is one of them.
• Freelancer.com category pages show open job counts and bid ranges live, so you can check demand before writing a proposal.

If your profile ranks for n8n today, add the terms buyers will search for next: dot setup, agent rules and approvals, agent security review, Gemini Enterprise agents and self-hosted automation.

n8n and Google's agent platform are in the same race

In May 2026 n8n's valuation doubled to $5.2 billion after SAP invested and said it would build n8n into Joule Studio. Its edge over Dots is control. You see every node, you can self-host it, and you choose the model.

Google made its move on April 22, 2026, at Cloud Next. Vertex AI became the Gemini Enterprise Agent Platform and Agentspace was folded into Gemini Enterprise. Workspace Studio arrived as a no-code way to build agents across Gmail, Docs, Sheets, Drive, Meet and Chat, with connectors for Asana, Jira, Mailchimp and Salesforce. Agent Designer, a visual flow canvas, is in preview, and version 1.2 of the Agent2Agent (A2A) protocol runs in production at 150 organizations.

For most clients the choice depends on where their data already lives. A company on Google Workspace will look at Workspace Studio first. A team that works in ChatGPT and Slack will try a dot. A company that needs audit trails and self-hosting will stay with n8n. A freelancer who can set up all three and explain the trade-offs is worth more than one who knows a single tool.

What full stack AI engineers can do this month

1. Learn how Dots handle rules and approvals, then sell dot setup and guardrails as a fixed-price service.
2. Rewrite portfolio pieces around results, such as hours saved, errors caught or money recovered, instead of the tool you used.
3. Learn to test computer-use agents. When an agent clicks through a web app, someone has to write the checks that catch a wrong click.
4. Keep one self-hosted stack, n8n with your own model keys, for clients who cannot use a cloud agent.
5. Price by outcome and maintenance, since the build step keeps getting cheaper.

What is still unclear

Before launch, OpenAI disclosed that its agents had posted 53 users' images online without authorization. Amazon has already blocked Muse from its store as an unauthorized agent, so dots may hit the same wall on some sites. Grok Bot has not published its task limits, how it stores credentials or which actions need approval. And every benchmark in this article comes from the company selling the model.

Frequently asked questions

What is OpenAI Dots?

Dots are always-on AI agents from OpenAI, launched on September 29, 2026. Each dot runs on its own cloud computer, uses more than 4,000 apps through plugins and can work for long periods without supervision.

Which model powers Dots?

GPT-6 Astra, released on September 22, 2026. Its API price is $10 per million input tokens and $50 per million output tokens.

How much do Dots cost?

One dot is included with ChatGPT Pro ($200 a month) and Business Premium ($125 a month). Enterprise, Edu and Healthcare workspaces can get a beta through their admin.

Does dot.com belong to OpenAI?

No. Registration records reported by Cybernews list Elon Musk as registrant, and the domain forwards to SpaceXAI's Grok Bot.

Is Dots better than Meta Muse?

They target different people. Muse is a consumer agent with a free tier. Dots are aimed at work and come with paid ChatGPT plans.

Will Dots replace n8n freelancers?

Simple, repeatable automations are at risk. Self-hosted setups, complex integrations, cost control and agent governance still need a person.

References

1. OpenAI, GPT-6 Astra: A new generation of intelligence
2. TechCrunch, OpenAI launches Dots, its bubbly agentic avatar
3. The Next Web, OpenAI launches dots, always-on AI agents with their own cloud computers
4. OfficeChai, OpenAI launched an AI agent named Dot, but dot.com redirects to rival Grok Bot
5. Cybernews, OpenAI's Dots launch runs into an awkward Grok problem
6. Trending Topics, Dot.com redirects to Grok Bot
7. Implicator, SpaceXAI Grok Bot starts at $120 a seat
8. Yahoo Finance, OpenAI debuts Dots AI agents in challenge to Meta's Muse
9. TechCrunch, Meta debuts its Muse AI agent
10. Fortune, Meta's Muse AI is exploding in popularity and drawing backlash from Amazon
11. DataCamp, GPT-6 Astra features, benchmarks and pricing
12. Upwork, In-Demand Skills 2026
13. Upwork, Hire AI agent developers (rates and reviews)
14. Fiverr, Business Trends Index 2026
15. Metaintro, Fiverr lost 21.9% of its buyers (Q2 2026 results summary)
16. Freelancer.com, n8n jobs
17. Freelancer.com, AI agents jobs
18. Press release, n8n valuation doubles to $5.2bn as SAP invests
19. The Next Web, Google Cloud Next 2026: AI agents, A2A and Workspace Studio
20. Wikipedia, Gemini Enterprise Agent Platform

Written by Zubair Hussain Shah. Published September 30, 2026. Figures are as reported by each source on that date and may change.