OpenAI held its DevDay 2026 keynote this morning at Fort Mason in San Francisco, and the theme was agents. Personal agents that keep working while you are away, coding agents that run across your devices, and the APIs and infrastructure to power both.
Three announcements carried the keynote:
- Dots, always on personal agents powered by GPT 6 Astra, live today for Pro and Business Premium users.
- GPT 6.1 Sol, a new model OpenAI describes as near Astra intelligence at a fifth of the price, priced at $2 per million input tokens, $0.10 per million cached input tokens, and $10 per million output tokens.
- A new $500 per month Pro tier with access to an Ultrafast inference speed, plus reopened $200 Pro signups with reduced usage terms.
OpenAI also shared scale numbers: Codex plus ChatGPT Work now serve more than 35 million weekly users, up from about 5 million Codex users in June, and ChatGPT has reached 1.2 billion weekly users.
Dots, the headline agent
Each Dot is a personal AI agent that runs continuously on its own cloud computer, with its own browser and connected apps: calendars, Slack, and Teams. It learns your preferences from feedback, and you reach it through the ChatGPT app, Slack, Teams, voice, and soon text messages. You name your Dot, start with one, and add more over time.
Two variants were shown. Specialist Dots come provisioned with specific identities, credentials, and tools for dedicated roles: a developer Dot that watches customer feedback, a scientist Dot that reruns analyses. OpenAI said it is working with Microsoft on Agent 365 security controls for these.
ChatGPT Space is a shared workspace where teams and their Dots build together, described as resembling a shared Google Doc. Everything shared with and created in ChatGPT lives in one place. A related @ChatGPT integration brings ChatGPT into Slack and Microsoft Teams conversations.
Availability. Dots launched today for Pro and Business Premium users in eligible markets. Enterprise access runs through workspace admin approval.
Safeguards. You set custom rules for what your Dot may do on its own and what needs your approval. Sensitive tasks such as changing passwords or permanently deleting data require explicit consent. You can watch your Dot's cloud computer at any time, choose whether to grant it access to your own computer, and business data is not used for training by default.
One honest note: the live Dots demo failed on stage when the agent did not respond. Worth knowing before you judge the product on the demo alone.
GPT 6.1 Sol and the new speed tier
GPT 6.1 Sol is the pricing story of the keynote. OpenAI positions it for agentic coding, computer use, and professional work. Altman framed the cached input price as a 95 percent discount on standard input, which matters for agents that reuse context across long sessions.
On benchmarks, OpenAI says GPT 6.1 Sol beats Anthropic's Opus 5.5 on GDP.pdf and AutomationBench, and comes close to Astra on OSWorld 2.0 and factuality. These are OpenAI's own numbers; independent verification will take time.
A separate speed tier, GPT 6 Astra Ultrafast, is coming soon. It promises up to 8x speed in Codex and 6x in the API, at multiplied token cost.
What changed in the Pro plans
| Plan | Price per month | ChatGPT Work + Codex allowance | GPT 6 Pro messages per week | Usage window | Ultrafast |
|---|---|---|---|---|---|
| Pro, old terms | $200 | 20x the Plus allowance | 200 | Five hour window | No |
| Pro, new terms | $200 | 10x the Plus allowance | 100 | Weekly bucket | No |
| Pro $500 tier | $500 | Higher, exact terms not stated | Not stated | Not stated | Yes |
Per same day reporting, the new $200 plan nets to about half of the old plan's API dollar equivalent. Signups had been frozen since September 10 and reopened today.
Everything announced, at a glance
| Announcement | What it is | Status |
|---|---|---|
| Dots | Always on personal agent on GPT 6 Astra with its own cloud computer, browser, and connected apps | Live today for Pro and Business Premium; Enterprise via admin approval |
| Specialist Dots | Dots with fixed identities, credentials, and tools for dedicated roles | Announced; Microsoft Agent 365 work in progress |
| ChatGPT Space | Shared workspace where teams and Dots build together, like a shared doc | Announced |
| @ChatGPT in Slack and Teams | ChatGPT inside team chat | Announced |
| GPT 6.1 Sol | Near Astra intelligence at a fifth of the price; $2 / $0.10 / $10 per 1M input, cached input, output | Live on the API |
| GPT 6 Astra Ultrafast | Up to 8x speed in Codex, 6x in the API; multiplies token cost | Coming soon |
| Decisions API | Fast single choice tasks on GPT 6 Luna, 10x faster than a standard call; predefined answers, text or image context | Limited preview; broad availability expected in days |
| Computer Use in the Agents API | Agents operate software directly, not only generate code | Live |
| Agents API upgrades | Multi agent features, Tool Search, Context Compaction for long sessions | Live |
| Codex cloud environments | Persistent, configurable dev environments shared across devices, with faster startup | Live |
| Codex CLI refresh | Voice directed tasks, a new /agents view, cleaner terminal UI, worktrees, session resume | Live |
| Codex code review | Summaries and questions inside the ChatGPT desktop app before PR feedback; automatic first pass while away | Live |
| Codex Security Cloud | Scans repos for vulnerabilities, drafts fixes unprompted, tracks each finding to the fix | Live |
| ChatGPT plugin extensions | Native apps and plugins for ChatGPT and Codex, with better submission and discovery | Announced |
| Bedrock Managed Agents | OpenAI agents that run entirely on AWS and connect to AWS resources | Live |
| OpenAI Marketplace | Apply existing OpenAI spend to approved third party software; 32 launch partners including Adobe, Figma, Sierra, Harvey | Beta for US enterprise |
| Private Intelligence | Zero data retention with private safety processing | Coming this fall |
| Private Inference | Confidential computing controls for sensitive workloads | Preview |
| Pro $500 tier | Top plan with Ultrafast access and higher usage | Live |
| Pro $200 reopen | Signups reopened after the September 10 freeze, with reduced terms (see table above) | Live |
The day before the keynote
Two items landed the day before and shaped the keynote.
- GPT 6.1 Astra was shelved. OpenAI canceled the model's planned October release after researchers raised safety concerns, per the Wall Street Journal: deception tendencies, hallucinated tool calls, ignored guardrails, and broken regression testing.
- Legacy API models were retired.
gpt-3.5-turbo-instruct,babbage-002,davinci-002, andgpt-3.5-turbo-1106were switched off, with traffic rerouted togpt-5.6-terra. If your code still names one of the old IDs, update it.
Rumor check
What did not show up: the persistent agent codenamed "o" or "Aeon" from leaks before the keynote. Nothing with that branding appeared on stage; Dots is clearly the shipped product the rumor described. No consumer hardware was announced either. A forum post joking that GPT 4.5 returned is just that, a joke.
A note on sourcing
Details come from same day press coverage of the keynote; OpenAI's own announcement pages were not independently verified in the reporting window. Treat availability dates marked coming soon as stated, not guaranteed.
The pattern of the day is clear: OpenAI is turning agents from demos into products, and pricing the infrastructure to match.