Lilith.
⌕
Editorial illustration: DevDay 2026: OpenAI turns ChatGPT into a workplace platform
Lilith illustration · editorial remix

OpenAI no longer wants ChatGPT to be the window where people occasionally ask a question. At DevDay 2026, it presented the service as the place where teams work, agents act and outside software reaches customers.

Dots keep working while Space gives teams a shared home

At Fort Mason in San Francisco on September 29, Sam Altman said ChatGPT now has 1.2 billion weekly users. The headline announcement was Dots, a new class of always-on agent powered by GPT-6 Astra. Each Dot receives its own cloud computer and browser, can connect to more than 4,000 applications and can carry responsibility for several projects instead of waiting for one prompt at a time.

People can talk or type to a Dot in ChatGPT, Slack and Microsoft Teams. SMS and phone access are on the roadmap. The agent keeps context as the conversation moves between channels, reports progress and asks questions when it needs a decision. Users can watch its virtual screen, and they can connect their own laptop when a task requires local access.

Dots and ChatGPT Space are available for Pro, Business Premium and Enterprise. Enterprise, Edu and Healthcare customers can enable an administrator-controlled beta, while OpenAI is also piloting Specialist Dots with their own identities, credentials and roles. The first Dot is included in a subscription. Conversation itself does not consume a ChatGPT allowance, but work handed to Codex or ChatGPT Work counts against those products' limits. More agents, greater speed and heavier workloads will cost extra. OpenAI is working with Microsoft on enterprise administration, though the precise product name was not clear enough in the keynote record to state with confidence.

Space gives people, ChatGPT and Dots a common knowledge base. Inside it, Pages supports collaborative writing, research, images, charts and other dynamic material. Space is available on the web and desktop for Pro, Business and Enterprise. Mobile users can search, read and share for now, with creation and editing promised later. Collaborative presentations and export to PowerPoint and Google Slides are due within weeks. OpenAI also announced team tasks, @ChatGPT in Slack and Teams, a beta meeting-notes plugin for macOS and shareable profiles.

The live demonstrations were useful precisely because they were not frictionless. During Holly Li's presentation, the voice conversation with Dotty stalled while the agent was still checking its work. A later handoff to a Codex thread also stopped before the team completed the demonstration by another route. Dots may be designed for continuous work, but the stage showed why continuous supervision still matters.

Sol lowers the price of intelligence as Codex moves into the cloud

GPT-6.1 Sol is the cheaper model at the center of OpenAI's new work stack. It is available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu customers, and through the API as gpt-6.1-sol. It is not available in the standard consumer chat. API pricing is $2 per million input tokens, $0.10 for cached input and $10 for output, one fifth of GPT-6 Astra's standard price.

OpenAI says Sol comes close to Astra in agentic coding, computer use and professional tasks. Compared with GPT-6 Sol, the company's own tests show gains of 6.4 points on DeepSWE, 4.8 on AutomationBench and 7 on OSWorld 2.0. The reported factual error rate fell from 11.4 percent to 7.7 percent, and the Terminal-Bench Science score more than doubled, although Astra still led at 68.1 percent. These are OpenAI's measurements, not results from an independent audit.

For customers willing to pay for latency, Astra Ultrafast promises up to 300 tokens per second. OpenAI says it can make Codex up to eight times faster and API output up to six times faster. It is available through the API and in ChatGPT Work and Codex for Pro 500 and selected Enterprise and Edu accounts, with a Sol version to follow. The speed carries a steep premium: $60 per million input tokens and $300 per million output tokens, six times Astra's standard $10 and $50 rates.

Codex now treats a development environment as a cloud project rather than something tied to one laptop. A developer can begin work on a phone, continue on a desktop or leave a long task running with the computer closed. Teams on Plus, Pro, Business, Healthcare, Education and Enterprise can share approved environments and permissions. The underlying Codex harness, which also powers Codex Work and Dots, is open source and can be embedded in other products.

The refreshed Codex CLI adds voice input through GPT-Live-1 and an /agents view, while the desktop application gains code review. The voice demonstration failed on stage, so Romain Huet switched to typed commands. Codex Security Cloud can inspect an entire GitHub repository on demand or on a schedule, follow new commits, consolidate duplicate findings and prepare a pull request with a proposed fix. OpenAI said it uses Daybreak Blue models.

The broader developer platform is moving in the same direction. Agents API now includes computer use, multi-agent coordination, tool discovery and parts of the Codex harness. Amazon Bedrock Managed Agents can run these capabilities alongside existing AWS applications and data. Decisions API, shown in a limited preview, uses GPT-6 Luna to choose quickly among predefined options from text or images; OpenAI did not announce a public release date or price. The company also said API time to first token has fallen by 45 percent and availability is above 99 percent. Damaged captions make any more precise claim about the number of nines unsafe.

Safety brakes remain visible, and the new Pro tiers cost more for less

OpenAI describes several boundaries around Dots. Background research is read-only, agents run in isolated environments unless a user connects a device, and Custom Rules can allow or block actions. Automated review is supposed to flag consequential steps. Password entry and other sensitive operations remain in human hands, and users can stop an agent at any time. OpenAI explicitly says fallible or high-stakes work requires human review. Business, Enterprise and Edu data is excluded from training by default.

Those controls matter because the release follows incidents involving autonomous behavior and a temporary pause for the most capable tool-using models. OpenAI withheld GPT-6.1 Astra after it failed to meet internal safety standards. Protesters outside Fort Mason raised a wider set of concerns, including job displacement, surveillance, data-center emissions and military use.

OpenAI also framed its models as participants in research. Tejal Patwardhan said an internal AI research intern can complete more than a third of day-long research tasks without intervention. Claims from the stage about progress on mathematics, antibiotics, historical languages, energy and robotics remain claims by the vendor, not independently verified findings.

One playful demonstration could easily be mistaken for a hardware launch. Lavender, a small robot, saw the audience, generated an image, responded by voice and moved in the physical world. It was a demonstration of real-time vision, voice, image generation and action through the same APIs, not a consumer device announcement. Altman promised future hardware worth waiting for, but announced none at DevDay.

At the same time, OpenAI reorganized its most expensive consumer subscriptions. Pro now comes in Pro 100, Pro 200 and Pro 500. The $500 monthly tier includes Ultrafast and an allowance described as 25 times Plus. New Pro 200 customers receive lower limits, while existing terms expire on October 29, 2026. CNET and Engadget report that Codex and ChatGPT Work fall from 20 times to 10 times the Plus allowance, and GPT-6 Pro falls from 200 to 100 messages a week. Those figures come from secondary coverage rather than independent testing for this article.

Sign-in, plugins and a marketplace make ChatGPT a distribution channel

The final group of announcements was less theatrical than Dots, but more revealing about OpenAI's business ambitions. Sign in with ChatGPT allows Plus and Pro subscribers to use part of their ChatGPT allowance inside partner applications, subject to weekly caps. Sixteen partners are participating at launch, including Cognition's Devin, Notion, Vercel, OpenClaw, T3, Dactyl and Nous Research's Hermes Agent. The authentication flow does not give those services access to a user's ChatGPT history or memory.

For large customers, the OpenAI B2B Marketplace turns existing contract commitments into a software budget. The keynote said more than 30 companies would be available at launch; secondary sources put the number at 32. Partners include Adobe, Figma, Salesforce and CrowdStrike. Baseten adds access to open-source and open-weight models through the marketplace, Codex and the Responses API.

Plugin extensions can now occupy a sidebar, open an editor or dashboard beside a conversation and provide custom viewers for files. The submission process also gains a simpler route to human review without forcing a developer to restart the application. The same plugins can work in ChatGPT Sites, which publishes shareable interactive pages and applications. Proposed MCP Events would let external events trigger plugin automations. Extensions are available across all plans.

Private Intelligence is OpenAI's attempt to make this expanding platform acceptable for sensitive work. Its preview combines Zero Data Retention with Private Safety Processing, so safety checks can run without storing customer content or exposing it to staff. Private Inference, developed with Cisco, Databricks and Snowflake around confidential computing, is planned for the fall.

Taken together, these releases are not just a larger feature list. OpenAI is assembling the workspace, the worker, the developer environment, the identity layer and the software store around one account. Whether customers embrace that convenience will depend on how reliably the agents behave, how clearly OpenAI prices their labor and whether its privacy safeguards hold up outside a keynote hall.

Lilith's verdict

OpenAI is building the office, staffing it with Dots and placing its own checkout at the door. The $500 floor is already open, but the safety inspection is still under way.