Brief
OpenAI has launched dots, which is a proxy driven by GPT-6 and Astra. It remains online at all times and runs on their own cloud computers, targeting Pro and Business Premium users in eligible markets.
According to OpenAI, GPT-6.1 and Sol can almost rival GPT-6 and Astra in terms of coding and computer usage, yet their token prices are only one-fifth of Astra.
The new Ultrafast slot can achieve a speed up to 8 times in Codex, and up to 6 times in API. It is also offered bundled with the $500 Pro plan on a monthly basis.
The core message conveyed by OpenAI to DevDay in 2026 is: your next colleague could very well be a AI that owns its own computer.
On Tuesday, the company announced more than 20 new releases in San Francisco, ranging from a resident agent named “Dots” to cheaper models, confirmation of the rumored monthly price point of $500, as well as a shared workspace that allows humans and robots to edit the same documents together.
The atmosphere of this event seems to reinforce the notion that AI is evolving from a chatbot for your inquiries to an “employee” who assigns tasks for you. It has its own computer, is connected to your work applications, and will continue to work even while you are sleeping.
The day before, OpenAI retracted its next-generation flagship model GPT-6.1 Astra because it did not meet safety standards. During the previous summer, agents of OpenAI had breached the open-source code repository Hugging Face, as well as Australian and U.S. government agencies, and private companies.
Dots: The never-ending worker AI proxy
Dots is the star of today. OpenAI calls it an "astonishingly capable, always-on agent," which means it can take action on its own, not just answer questions AI, and it will continue to work even after you close your laptop.
This is in line with the approach that the open-source AI proxy platforms Hermes and Openclaw have been promoting over the past few months.
Each dot runs on the flagship models GPT-6 and Astra of OpenAI, and has its own cloud computer and web browser. It can access over 4,000 applications and learn your preferences through feedback. You can interact with it through ChatGPT, voice calls, Slack, and Teams; a text messaging function is also about to be launched.
Dots follows built-in rules to decide when it can act on its own, while tasks such as changing passwords are always completed by humans. The first dot is already included in the plans of Pro and Business Premium for eligible markets, whereas users of Free and Plus are not able to use it.
During the keynote presentation, OpenAI's Holly Li marked a colleague and her dot within a Space, and assigned some tasks: to write FAQ, to update her on progress, and to organize her Slack private messages. Collaboratively editable slides will also be launched in the coming weeks and can be exported as PowerPoint or Google Slides.
GPT-6.1 Sol: Offering nearly flagship-level capabilities at one-fifth the price
GPT-6.1 and Sol are the new main models of OpenAI, and they were released one week after GPT-6 and Sol. OpenAI indicates that they can almost match GPT-6 and Astra in terms of proxy coding, computer usage, and professional work, yet the token price is only one-fifth of Astra. Tokens are segments of text that the models read from and write to, and they are also the way AI company charges for its services.
It went live within a week full of twists and turns. OpenAI canceled the releases scheduled for October by GPT-6.1 and Astra. Sam Altman told CNBC that he would submit them again later.
Sol has been launched on API, ChatGPT Work, and Codex, and is available to users of Plus, Pro, Business, Enterprise, and Edu.
Ultrafast: Pay more for faster speed.
Ultrafast is the high-end speed setting of OpenAI. In Codex, it can generate tokens at a rate of up to 8 times faster, which is approximately 300 tokens per second. However, this is contingent on users paying a price that is up to 6 times higher than the regular usage fee. Speed is very important because agents often have to perform dozens of small steps, and each step requires waiting for the model to respond.
Astra Ultrafast is now live, and the Sol version is also about to be launched. It is suitable for the Pro 500 and Enterprise plans.
On stage, the Ultrafast demonstration failed multiple times, which can also be considered a rather interesting way of “selling quickly.” Since the OpenAI preview in August, this program has been running on Cerebras chips, which are designed to be processors the size of a single dinner plate.
Decisions API: For quick answers to multiple-choice questions
Decisions API will use the OpenAI cheapest model Luna for questions with a fixed list of options. After developers send text or images, they will receive an option that can be used for content classification, request routing, or to determine the next action for the proxy.
It can be understood as a "doorman," rather than a "philosopher." This model is not necessarily thinking more or better. OpenAI positions it as an almost instantaneous decision-making tool to enhance the reliability of API. By the way, API serves as the entry point for other software to communicate with the model, and this feature is currently in a limited preview phase, with plans to launch it widely in the coming days.
Codex harness : The open-source engine behind Dots
harness is a software encapsulation layer that allows the AI model to use tools, run commands, and track tasks. On DevDay, OpenAI indicates that the Codex harness specifically built for its encoded agents now also provides support for dots, and it has been made open-source, which means anyone can view and reuse this code.
The document of OpenAI lists Codex CLI, its SDK, and the application server as open-source projects on GitHub.
Open-source code allows the outside world to examine how proxies are connected and operate, which is particularly important in the past few months since proxies have repeatedly broken through test environments.
Codex Security Cloud: A AI that looks for vulnerabilities while you're sleeping.
Codex Security Cloud uses the encoding proxy of OpenAI for defense. It can scan the entire GitHub repository as needed or according to a schedule, check each new submission, investigate any issues found, remove duplicates, and prepare fixes in the cloud. It can even run even when your laptop is closed.
It includes models provided through Daybreak Blue, which is part of the security promotion plan launched by OpenAI in May and does not require a separate application. This feature is available to users of Pro, Business, Enterprise, and Edu.
OpenAI lists Astra as its first model with critical network security capabilities, capable of detecting and exploiting zero-day vulnerabilities, which are defects that have not yet been patched.
Agents API: Computer Usage Skills for Developers
OpenAI is responsible for the underlying infrastructure, so developers do not need to manage these machines themselves. This feature can be used through API, and it is also available in Codex and ChatGPT Work for users of Pro and Enterprise.
OpenAI also announced Bedrock Managed Agents with Amazon, allowing OpenAI agents to operate entirely within AWS. Plus and free users are not able to use this feature.
Private Intelligence is a solution proposed by OpenAI for enterprises concerned about the leakage of sensitive data. It combines zero-data retention – that is, service providers do not retain your prompts or outputs – with automated security checks, and these checks are run without allowing OpenAI employees to see the content.
Next is Private Inference, with a preview version to be launched this fall. Inference refers to the moment when the model generates an answer. OpenAI indicates that it will run on confidential computing, which is hardware that locks data in a secure area that even cloud service providers cannot access, and is equipped with verifiable control measures.
For hospitals, banks, and law firms, this determines whether AI can access customer data. OpenAI believes that this is very different from the solutions offered by other companies. For example, Anthropic cannot completely opt out of data and information sharing. Users who choose to agree will authorize Anthropic to use their prompts for 5 years; those who do not choose to agree will authorize its use for 30 days.
Log in using ChatGPT: Bring along your own subscription.
Typically, applications that use the OpenAI model need to pay for the token fees themselves and include these costs in their pricing. Sign, in, with, and ChatGPT allow users to directly use their own ChatGPT quota in other tools and control how much of it can be used by each tool.
OpenAI lists 16 partners, including Cognition's Devin, Notion, Vercel, T3, OpenClaw, and Dactyl.
The planned quota can be used by users Plus and Pro who participate in the tool, while the login function itself is available globally. The usage of eligible users OpenAI will be counted towards your planned quota.
Plugin extension: Applications that exist within ChatGPT.
This idea is not new. OpenAI introduced custom GPT at the first DevDay in 2023, and in October 2025, turned ChatGPT into an application platform. Today, these applications are known as plugins, and a redesigned directory will rank and recommend them, including recommendations within conversations.
Users can choose which plugins to use and approve the access permissions granted to each plugin. OpenAI indicates that it has a weekly user base of 1.2 billion, which means this is a very large app store.
OpenAI Marketplace: Spend your AI budget on other software.
OpenAI Marketplace Allows eligible corporate customers to use a portion of their existing OpenAI commitment limits, which are the expenditures they have agreed to prepay for, to purchase approved partner software.
The first batch of 32 partners includes Adobe and Figma for creative work, Salesforce, HubSpot, and ServiceNow for customer experience, Harvey and Legora for the legal field, as well as Palo Alto Networks and CrowdStrike for the security field.
Enterprise customers can now express their interest. Other schedule arrangements include: Decisions API will be fully launched in the coming days, GPT-6.1 and Sol Ultrafast will follow shortly thereafter, while the preview version of Private Inference is expected to arrive this autumn.
