DevDay 2026: OpenAI builds an entire work stack around Dots

At DevDay 2026, OpenAI announced more than twenty components of a ChatGPT work platform.

Dots get their own computers while teams get Space

At Fort Mason in San Francisco on September 29, Sam Altman cited 1.2 billion weekly ChatGPT users. Dots are GPT-6 Astra agents. Each gets a cloud computer, browser and connections to over 4,000 apps. Users reach them across ChatGPT, voice, Slack and Teams, with SMS planned. Context persists; users can inspect screens or connect a laptop.

Dots are rolling out to Pro and Business Premium in eligible markets; Enterprise, Edu and Healthcare get an admin beta. One is included. Conversation is allowance-free; Codex and ChatGPT Work tasks count, while extra dots and compute cost more. Enterprise pilots add identities, credentials, roles and Microsoft Agent 365 integration.

ChatGPT Space puts teams, ChatGPT and dots on one knowledge base. Pages support writing, research, images and charts. Space is on web and desktop for Pro, Business and Enterprise; mobile supports discovery, reading and sharing, with editing later. PowerPoint and Google Slides export is due within weeks. OpenAI also announced shared tasks, @ChatGPT in Slack and Teams, a beta macOS meeting-notes plugin and shared profiles. Holly Li's live Dots voice demo stalled, as did a later handoff into a Codex thread; both demonstrations continued by another route.

Sol cuts model costs, Ultrafast sells speed and Codex moves to the cloud

GPT-6.1 Sol is available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu, and through the gpt-6.1-sol API, but not standard ChatGPT Chat. It costs $2 per million input tokens, $0.10 for cached input and $10 for output, one fifth of GPT-6 Astra. OpenAI claims near-Astra agentic coding, computer use and professional work. Against GPT-6 Sol, it reports gains of 6.4 points on DeepSWE, 4.8 on AutomationBench and 7 on OSWorld 2.0; factual errors fell from 11.4% to 7.7%. Terminal-Bench Science more than doubled, though Astra leads at 68.1%. None has an independent audit.

GPT-6 Astra gains Ultrafast in the API, ChatGPT Work and Codex for Pro 500 and selected Enterprise and Edu accounts. OpenAI claims up to 8x faster Codex at 300 tokens per second and 6x faster API output; Sol follows later. Pricing of $60 per million input tokens and $300 per million output tokens is six times the standard $10 and $50.

Codex offers cloud projects and approved environments for Plus, Pro, Business, Healthcare, Education and Enterprise. Its open-source harness powers Codex Work and Dots and can be embedded in developers' own products. The CLI adds GPT-Live-1 voice and /agents; desktop ChatGPT adds code review. The live CLI voice demo failed, so the presenter switched to the command line. Codex Security Cloud scans GitHub repositories and prepares fixes with Daybreak Blue. Agents API adds computer use and multi-agent coordination. The limited Decisions API uses GPT-6 Luna to classify and route fixed choices. AWS Bedrock Managed Agents gains core Agents API capabilities. OpenAI also said API time to first token fell 45% and availability was above 99%; the caption record does not support a more precise number of nines.

Safety brakes collide with a smaller and more expensive Pro

Dots research is read-only, and agents remain isolated unless a device is connected. Custom Rules govern actions; automated review checks consequential steps, while passwords and sensitive operations remain human-only. Users can stop agents. OpenAI requires human review for fallible, high-stakes work. Business, Enterprise and Edu data is excluded from training by default.

OpenAI says its internal AI research intern completes more than one third of day-long internal research tasks without intervention. Its related stage claims about progress in mathematics, science, antibiotics and robotics are vendor claims, not independently verified results.

Dots arrive days after autonomous-behavior incidents and a temporary pause of the most capable tool-using models. GPT-6.1 Astra was withheld after missing internal safety standards. Protesters outside Fort Mason cited labor disruption, surveillance, data-center emissions and military use. No consumer hardware launched. Lavender was a multimodal robot demo combining vision, voice, image generation and physical movement, not a consumer product announcement. Altman promised something worth waiting for.

Pro splits into Pro 100, Pro 200 and Pro 500. The $500 plan includes 25x Plus and Ultrafast. Pro 200 limits shrink; legacy terms end October 29, 2026. CNET and Engadget, not independent editorial tests, report Codex and ChatGPT Work falling from 20x to 10x Plus and GPT-6 Pro from 200 to 100 weekly messages.

Login and marketplace turn ChatGPT into a distribution layer

Global Sign in with ChatGPT lets Plus and Pro users spend allowances across 16 apps with weekly caps. Partners include Cognition's Devin, Notion, Vercel, OpenClaw, T3, Dactyl and Nous Research's Hermes Agent. Authentication exposes neither history nor memory. OpenAI Marketplace lets enterprises apply commitments to products from 32 partners including Adobe, Figma, Salesforce and CrowdStrike. Baseten brings open-model inference to Marketplace, Codex and the Responses API.

Plugin extensions gain sidebar placement, canvases, file viewers and simpler review. Sites publishes shareable interactive pages and apps linked to plugins; proposed MCP Events would trigger plugin automations from events. Extensions are available on every plan.

Private Intelligence previews Zero Data Retention with Private Safety Processing, checking safety without retaining data or staff access. Private Inference, based on confidential computing and developed with Cisco, Databricks and Snowflake, is planned for fall. Adoption and safeguards in daily use will test this tightly bound stack.

Lilith's verdict

OpenAI has built an office complex around its customers, and Dots already carry the access badge. Now safety must keep pace with the elevator to the $500 floor.