On September 29, OpenAI showcased over 20 new products and updates during its annual DevDay conference. The highlights included the introduction of Dots, persistent AI agents, the new GPT-6.1 Sol model, enhancements to Codex and Agents API, and collaborative tools within ChatGPT.
The company reported that its chatbot now reaches a weekly audience of 1.2 billion users.
Dots
Dots are AI agents built on the GPT-6 Astra framework, designed to handle long-term tasks. They retain context between sessions and can work autonomously without constant user involvement.
Each Dot operates on its cloud computer and browser, allowing connections to over 4,000 applications via plugins. Users can interact with Dots through platforms like ChatGPT, Slack, and Microsoft Teams.
For instance, a Dot can monitor a work project, gather new information from connected services, and return with a completed result.
Background access to applications is limited to reading data. Actions that modify data, affect accounts, or share information with third parties may require separate permissions and user confirmation.
Currently, users can create only one primary Dot. This feature is available on Pro and Business Premium plans in supported regions, while an opt-in beta version has been launched for Enterprise, Education, and Healthcare users.
GPT-6.1 Sol
This model is primarily geared towards programming, computer management, and professional tasks.
According to OpenAI, in various scenarios, its capabilities are comparable to those of GPT-6 Astra, while the cost for input and output tokens for GPT-6.1 Sol is five times lower than that of Astra.
The API pricing is set at $2 per million input tokens and $10 per million output tokens. Cached input costs $0.10 per million tokens, with standard input and output prices remaining the same as GPT-6 Sol, and cached input prices halving.
In the DeepSWE v1.1 benchmark, which evaluates AI agents on long-term programming tasks in real codebases, GPT-6.1 Sol matched Astra while costing about one-fifth of the task's execution cost.
It also outperformed GPT-6 Sol by 6.4 percentage points.
In OSWorld 2.0, the new model lagged behind Astra by 2.1 percentage points at maximum reasoning levels, but completing a task cost approximately seven times less.
OpenAI also assessed the actual accuracy on a set of complex queries. With low reasoning depth, the percentage of answers containing at least one error dropped from 11.4% to 7.7%.
OpenAI emphasized that this sample consisted of problematic queries and does not reflect typical model usage.
GPT-6.1 Sol is accessible through API, ChatGPT Work, and Codex for Plus, Pro, Business, Enterprise, and Education users, but it is currently unavailable in standard ChatGPT chats.
In the coming days, OpenAI plans to introduce an Ultrafast mode for GPT-6.1 Sol. This mode is said to enable up to eight times faster token generation in Codex compared to the standard setting, reaching up to 300 tokens per second. In the API, the acceleration can be up to six times. Ultrafast is currently available for GPT-6 Astra, with support for GPT-6.1 Sol to follow later.
Initial Reactions
Within hours of the release, early feedback on GPT-6.1 Sol appeared on Reddit. Some developers noted low limit consumption and significant improvements over GPT-6 Sol.
Users reported that after several hours of use, they barely touched their weekly limit, though they do not see Sol as a complete replacement for Astra in all scenarios.
Negative reviews surfaced as well, with some users complaining about errors in terminal operations, following instructions, and tool management. Others noted that the model operates slower than Astra at comparable reasoning depths.
Feedback on Dots has been less frequent, with discussions primarily focusing on the feature's availability, privacy concerns, and user willingness to grant a persistent agent access to emails, files, and other services.
ChatGPT Space
OpenAI also introduced ChatGPT Space, an environment where employees, ChatGPT, and Dots can collaboratively work on shared project contexts.
This space includes Pages—editable documents for texts, research, charts, and images. Users can invite colleagues to collaborate in real-time with AI.
In the coming weeks, the company promises to add collaborative presentation creation capabilities. Multiple users and agents will be able to edit slides simultaneously, leave comments, and export results to PowerPoint or Google Slides.
For Business and Enterprise users, team tasks are now available, which can be scheduled or triggered by specific events, such as receiving an email or a new message in Slack.
Codex
Codex now supports running tasks in the cloud, allowing users to return to them from different devices, even if the primary computer is turned off.
The CLI has introduced voice control and a /agents section for monitoring multiple tasks simultaneously.
A separate Code Review interface has been added to the desktop application, displaying changes in files and enabling discussions with Codex before sending comments to GitHub or GitLab.
Codex Security Cloud is designed to audit GitHub repositories. The service builds a threat model for the codebase, scans the change history, checks for potential vulnerabilities in an isolated environment, and suggests fixes for review.
The Agents API has been enhanced to include computer control, multi-agent collaboration, tool searching, and automatic context compression. OpenAI manages sessions and orchestration, while developers define the tools and execution environments available to the agent.
OpenAI also introduced the Decisions API, aimed at tasks with a limited set of options, such as content classification, request routing, or determining the next action for an agent.
Plugins
Plugin developers will now be able to add dedicated sections to the ChatGPT sidebar, create interactive interfaces, and build custom file viewers. OpenAI has also updated its tools for creating and publishing plugins.
The company has incorporated support for MCP Events, a specification that allows for automatic actions triggered by events in connected services.
On September 29, Bloomberg sources reported that OpenAI is in discussions to raise $30 billion with a valuation of $1.4 trillion.
Stay updated by following ForkLog on social media:
Telegram (main channel) Facebook X