Dots and GPT-6.1 Sol: what OpenAI's DevDay changes for businesses

What OpenAI announced on September 29

At its DevDay on September 29, 2026, OpenAI made two announcements that directly concern businesses. The first is a product: dots, always-on agents that work in the background for their user. The second is a model: GPT-6.1 Sol, an improved version of GPT-6 Sol that comes close to GPT-6 Astra, OpenAI's most capable model, at a much lower price.

OpenAI also launched faster versions of its models, GPT-6 Astra Ultrafast and GPT-6.1 Sol Ultrafast, available in the API and to a new Pro tier at 500 dollars per month.

These announcements come a week after the release of Claude Opus 5.5 and GPT-6 Sol, and the day after Claude Sonnet 5.5. The pace is fast, and every announcement comes with flattering test results. We read OpenAI's official pages, including the help page on dots, which contains details missing from the announcement. Here is what they say, and what a European business should take from them.

Dots: a permanent agent with its own computer

OpenAI presents dots as "always-on" agents powered by GPT-6 Astra. Each dot has its own cloud computer and its own browser. It can write and test code, connect to more than 4,000 apps through OpenAI's plugin ecosystem, and work on its user's goals around the clock.

The difference from a regular conversation is continuity. A dot keeps context from one request to the next and across channels: ChatGPT on web, desktop or mobile, but also Slack, Teams and, in some countries, text messages. It learns from your feedback, forms memories, including from connected apps, and can take on several projects in parallel.

Above all, it does not always wait to be spoken to. OpenAI describes "proactive research": when you are not working with it, the dot goes through connected apps looking for ways to help. For this background research, it uses restricted tools that cannot send messages, change app content, or control your browser or computer.

OpenAI's examples range from fixing bugs reported by customers to updating a sales proposal during a technical review, or refreshing an analysis when new data arrives. In one case cited, a tester's dot noticed a forgotten invoice, prepared it, and sent it after its user approved.

In Europe: not on Pro, but on Business Premium

This is the most useful information for a European business, and it is not in the main announcement. OpenAI's help page states that dots are rolling out to Pro users in markets excluding the European Economic Area, Switzerland and the UK. A Pro account in France therefore has no access for now.

Dots are, however, available to Business Premium users in all regions supported by ChatGPT. For Enterprise accounts, a beta exists but is turned off by default: the workspace admin must enable it.

Two other details matter. The rollout is gradual and access may take several days. And for the first month, dots usage will not count toward the allowances of eligible plans; OpenAI says it will then share usage terms for each plan. In other words, the real usage cost is not known yet. It is wise not to build a critical process on a tool whose terms will change within a month.

The controls provided, and what they imply

OpenAI details several control mechanisms. They are serious, but they leave the business with decisions it must make explicitly.

• Action rules. For some actions, such as sharing, purchasing or accessing something, the user chooses between four behaviors: act without asking, act if pre-approved in the request, ask before acting, or hand off to the user. Some sensitive tasks, such as changing a password, are always reserved for the human.
• Automatic checks. Actions that may affect your accounts or share information are checked against your instructions, your rules and OpenAI's safety requirements.
• A separate computer. The dot works on its own cloud computer. Access to your local computer is optional and off by default.
• An activity view and a pause. You can follow ongoing work, including background work, and pause the dot at any time.

One point deserves particular attention. According to the help page, disconnecting an app does not delete the information the dot has already obtained from it. To remove it, you must reset the dot, which also deletes its conversations, memories and scheduled tasks. Connecting a mailbox, a CRM or a document space to a dot is therefore not a trivial, reversible step. It is a data governance decision, to be made before connecting, not after.

OpenAI itself reminds users that dots can make mistakes, including when following your rules, and that high-stakes work must be checked. It is the same logic we apply: what commits the company is never left to the model alone, as explained in why an AI agent must never quote a price.

Specialized dots: an agent with a role and an identity

The most interesting part for organizations may be the least visible. OpenAI is offering a limited number of companies a preview of specialized dots: agents configured by the company with their own identity, credentials and access to the systems needed for a specific responsibility. OpenAI cites its first internal trials in procurement, invoice processing, email marketing, customer support and sales contracts.

Two details of the announcement are telling. First, OpenAI says its engineering teams will work directly with organizations to define each dot's responsibilities, its tools and how its work is checked and approved. That is the principle of the engineer deployed at the client, which we describe in what is a forward deployed engineer. Even with a very capable agent, value is built in defining the role, the access and the controls.

Second, OpenAI is working with Microsoft to integrate specialized dots with Agent 365 governance and security controls and with Microsoft's identity controls. The stated goal is for IT teams to manage dots' access and oversee their activity with the Microsoft tools they already know. For a company on Microsoft 365, this echoes the question we raised about Claude and GPT-6 Sol arriving in Copilot: who decides access, for which data, and who is accountable.

GPT-6.1 Sol: close to Astra, at a fifth of the price

The second announcement concerns teams building their own tools. GPT-6.1 Sol is presented as a clear improvement over GPT-6 Sol in coding, computer use and professional work. OpenAI positions it as close as possible to GPT-6 Astra, at a fifth of Astra's standard prices.

Its API prices are 2 dollars per million input tokens, 10 dollars output, and 0.10 dollars for cached input tokens, 95% less than the standard input price. That last point matters for agents, which often reuse the same context from one request to the next.

On results, OpenAI highlights several tests. On AutomationBench, which measures execution of multi-step business workflows, GPT-6.1 Sol beats Claude Opus 5.5 by 2.2 points at medium effort, for about a third of the cost. On a factuality evaluation built from conversations where errors had been flagged, its error rate is 4.1%, against 4.5% for GPT-6 Sol and 4.0% for Astra at the same setting.

In parallel, OpenAI launched GPT-6 Astra Ultrafast, presented as up to 8 times faster than GPT-6 Astra, and GPT-6.1 Sol Ultrafast. According to OpenAI, this fast version of Sol offers near-Astra intelligence at up to 8 times the speed, for an API cost roughly equivalent to what GPT-6 Astra cost before. Speed comes at a price: for an agent answering customers live, it may be worth it; for an overnight batch job, it rarely is. The choice depends on how long your users can reasonably wait for an answer.

The model is available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu subscribers, and in the API as gpt-6.1-sol. OpenAI notes that it is not yet available in Chat.

Reading these announcements without being misled

Three precautions apply to both announcements.

The numbers are the vendor's. OpenAI states that its evaluations were run in its research environment or through the API, with results that may differ from ChatGPT in production, and that competitor scores come from public reports. Comparisons with Claude Opus 5.5 are therefore comparisons chosen by a competitor. Anthropic does exactly the same in its own announcements, as we saw with Claude Sonnet 5.5.

The rollout is gradual. OpenAI's help page states that some features announced at DevDay are not yet available to all accounts. What was shown on stage is not necessarily what your teams will see this week.

Terms are not set. Free dots usage for a month, future subscriptions to add dots or speed up their work, the 500 dollar Pro tier: the business model is still being built. To manage these costs once they are known, our CADRE framework tracks them per use and per agent.

What it changes for your projects

Dots make visible a distinction we draw in all our projects: the personal agent and the business agent do not serve the same need.

A personal agent, like a dot, serves a person. It knows their preferences, apps and priorities. It is very useful for individual work, and its mistakes are caught by its user, who reviews what it produces.

A business agent serves a process. It handles invoices, answers customer requests or qualifies files, on behalf of the company rather than a person. It must behave the same way every time, rely on written rules, leave a verifiable trail and meet regulatory obligations. That is exactly what OpenAI is piloting with specialized dots, with its own engineers, and what we build with our clients, as described in deploying an AI agent in your company, step by step.

Before connecting a dot, or any permanent agent, to company tools, a few decisions are required: which apps it may read, which it may change, which actions require approval, which data must never be entrusted to it, and how to erase what it has learned if you stop. These questions overlap with the AI Act and the GDPR, which we detail in AI Act: what already applies to your SMB in 2026.

Our takeaway

Dots mark a step: the assistant that answers becomes an agent that acts continuously, with its own access, memory and initiative. The promise is real, and the controls OpenAI provides are serious. But the value, like the risk, lies in the decisions a company makes before plugging the agent into its tools.

GPT-6.1 Sol confirms the trend of recent days: models close to the top level, at a fraction of the price. For the tools you build, that is good news, provided you check it on your own cases, not on a ranking published by the vendor.

If you are wondering which work to give a personal agent, which deserves a business agent, and which rules to set before connecting anything, our AI diagnostic starts from your real processes and data constraints.

Sources

• OpenAI, Introducing dots, September 29, 2026 (French edition)
• OpenAI Help Center, Getting started with your dot
• OpenAI, Introducing GPT-6.1 Sol (French edition)

Frequently asked questions

Can we use dots in Europe today?

Not with a Pro account: OpenAI's help page excludes the European Economic Area, Switzerland and the UK for that plan. Business Premium accounts have access in all supported regions, and Enterprise accounts can try the beta if their admin enables it.

How much do dots cost?

The first dot is included in Pro and Business Premium plans, and its usage is not counted for the first month. OpenAI then announces usage terms per plan and paid options to add dots or speed up their work, without publishing prices so far.

Can a dot act without our approval?

Yes, for actions you allow it to take alone. Custom rules let you require approval or block certain actions, and a few sensitive tasks always stay with the human. These rules must therefore be set before connecting apps.

What happens to what a dot has learned if we disconnect an app?

According to OpenAI, information already obtained is not deleted by disconnecting. To remove it, you must reset the dot, which also deletes its conversations, memories and tasks. This belongs in your data policy.

Is GPT-6.1 Sol better than Claude for our agents?

The published comparisons are OpenAI's, on tests it chose. The only way to decide is to run your own cases on the candidate models, at the same effort level, measuring quality, cost and delay per task.