Recap of OpenAI DevDay 2026: Personal Agent, Sol 6.1, and the Absent Astra

On September 29, 2026, in San Francisco, despite the model competition being stifled by old rival Anthropic, OpenAI DevDay 2026 proceeded as scheduled.

However, the new flagship model was still absent from this launch event.

The originally planned October release of GPT-6.1 Astra was urgently pulled by officials just twenty-four hours before the opening. The new protagonist stepping into the spotlight became the Personal Agent, named Dot by OpenAI.

This launch unveiled 25 updates, summarized by Dongcha Beating below.

Models and Dot

Dot

A Personal Agent powered by GPT-6 Astra. It features an independent cloud sandbox computer and a dedicated browser, integrating with over 4,000 external applications; it supports proactive research, capable of browsing apps for tasks in the background with read-only permissions; Pro and Business Premium users receive their first dot free of charge, and conversations do not count against the ChatGPT base quota.

Specialist Dots

A preview version of digital employees for enterprises. They possess independent identities, credentials, and employee IDs within the organization.

GPT-6.1 Sol

Focused on agent programming, computer operations, and professional tasks. Input costs $2 per million tokens, output $10 (one-fifth of Astra's), cached input $0.10; DeepSWE v1.1 scores match Astra, while Terminal-Bench scores exceed the previous generation Sol by more than double.

Astra Ultrafast & Sol Ultrafast

High-throughput ultra-fast tier. Speed increases up to 8x, with Codex throughput at approximately 300 tokens per second. Priced at 6 times that of standard Astra (input $60, output $300).

Privacy Intelligence

Full series supports 'zero data retention' with offline security reviews; jointly launched privacy inference based on confidential computing architecture with Cisco, Databricks, and Snowflake.

Codex and API

Codex Cloud

Fully managed cloud development environment. Supports offline operation; processes continue when the laptop is closed, allowing teams to share pre-configured containers.

New Codex CLI

Redesigned terminal interaction, supporting real-time voice commands, native worktree support, and a new /agents view for multi-agent scheduling and progress tracking.

Code Review

Integrated into the desktop app, supporting GitHub PRs and GitLab MRs, capable of automatically performing initial reviews and diff diagnostics via the cloud in offline mode.

Codex Security Cloud

Productized internal security project Defense Factory, integrated with the Daybreak Blue cybersecurity large model, providing continuous vulnerability scanning and automatic remediation for code repositories.

Decisions API

OpenAI's version of Jev, a specialized decision interface locking GPT-6 Luna within fixed Q&A scopes. Dedicated to intent classification and agent path routing, with response times around 150 ms, offering a 10x speed increase over conventional calls.

Agents API Opens Computer Use

Fully opens the core capabilities of the Codex harness (multi-agent orchestration, context compression, tool retrieval, and screen operations). Framework code is open-sourced, with full management also provided.

Amazon Bedrock Managed Agents

Deeply integrated with AWS, allowing enterprises to directly run OpenAI's agent systems within their own AWS VPCs and compute resources.

Plugin Ecosystem

Plugin Sidebar and File Interaction

Developers can customize exclusive interaction panels or file viewers in the ChatGPT sidebar, directly embedding products into the conversation flow.

Directory Refactoring and Passive Discovery

Revamped plugin submission and review flows, proactively recommending relevant plugins during natural user conversations based on context.

Plugin Access Sites

Supports enterprises building their own sites to directly connect plugins, with members connecting to internal enterprise data according to their respective permissions.

MCP Event-Driven

Integrates the MCP event specification advocated by Anthropic; when external collaboration tools (like project boards) undergo status changes, plugins are automatically awakened in the background to draft proposals.

Collaborative Office

ChatGPT Workspace

Replaces the original library, serving as a long-term collaborative hub where team members, Codex, and dots share the same context.

Dynamic Pages

Rich text carriers co-edited by multiple people and agents, supporting dynamic components, real-time to-dos, and data dashboards, allowing targeted @ mentions to agents for modifications on selected text.

Collaborative Slides

A presentation engine supporting real-time multi-user collaboration, enabling generation and demonstration within ChatGPT, with lossless export to PowerPoint or Google Slides.

Team Tasks and Scheduled Dispatch

Creates periodic task flows within enterprises that can be triggered by email, IM, or schedules.

Office IM Integration

Deeply embedded in Slack and Teams, allowing enterprise members to @ invoke summaries or bug troubleshooting in channels without needing personal accounts.

Meeting Plugin

Debuting on macOS, locally recording audio and generating key decisions and follow-up to-dos combined with context; audio is destroyed immediately after offline analysis.

Personal and Team Skill Profiles

Supports centralized display and reuse of personally built sites, plugins, and shared skills.

Commercialization and Subscriptions

Sign in with ChatGPT

Unified identity authentication across the web. Plus and Pro users can directly consume their subscribed token quotas within third-party products; the first batch includes 16 companies such as Devin, Notion, and Vercel.

New $500 Pro Tier

Provides the highest priority concurrent quota, exclusively accessing Astra Ultrafast.

Adjustment of Original $200 Pro Quota

Reopened after a 20-day pause, but starting October 30, the compute quota for Codex and ChatGPT work shrank from 20x that of Plus to 10x, and the weekly cap for GPT-6 Pro was halved to 100 requests. Existing users received a $2,500 credit expiring before year-end.

OpenAI Software Marketplace

In partnership with 32 software vendors including Adobe, Figma, Salesforce, ServiceNow, and Harvey, enterprises can use pre-committed consumption quotas for OpenAI to offset and purchase third-party SaaS products.

The only true product of this DevDay is Dot.

The other releases, whether Sol, Ultrafast, Codex harness, Space, Pages, or MCP events, are all scattered parts when viewed individually, piecing together to form the flesh, bones, and nerves of Dot.

Dot thinks with Astra, operates the system with Codex harness, connects to over 4,000 external software via plugins, wakes itself up in the background when board statuses change via MCP events, writes completed tasks into Pages, and posts them to Teams channels.

For the past three years, human-machine interaction 默契 has remained confined within that small box.

First, it was 'you ask one question, it answers one'; then it became 'you assign a task, it runs and delivers'.

Dot no longer waits for you to hit Enter at the cursor. It possesses its own cloud computer and browser, continuing to browse various apps with restricted read-only permissions in the background even after the user closes the screen, looking for whose invoice hasn't been issued or which bug remains unpatched.

Officials say it will complete the work according to your workflow before you even speak.

In the past, whether ChatGPT or various Copilots, the 争夺 was for software-dimensional seat licenses, charging monthly rent per head, essentially equipping employees with a handy screwdriver.

Dot and Specialist Dots go further. Specialist Dots possess independent domain accounts, system credentials, and employee IDs within enterprises, integrating into Microsoft's Agent 365 governance system, following IT departments' familiar access, audit, and deactivation processes.

Enterprises are no longer buying a tool account; they are hiring a machine employee who doesn't require social security contributions and never goes offline.

This is where the two Silicon Valley giants diverge. Meta pushes the similarly shaped Muse to billions of ordinary people, leveraging its #1 free chart position to scale consumer numbers; OpenAI, however, seals Dot behind the high walls of Pro and enterprise subscriptions, with more Dots requiring monthly add-ons in the future.

One competes for users' time, the other for corporate workstations.

Acceleration

On stage are machine employees, but behind the scenes lies a compute bill filled with anxiety.

Pricing for GPT-6.1 Sol is only one-fifth of Astra's, yet performance closely tracks it on most benchmarks; Astra Ultrafast lists a standard price 6 times higher, exchanging for ultra-fast speeds of 300 tokens per second; immediately following, Sol Ultrafast is packaged as 'buying brains close to Astra and 8x speed for the original price of Astra'.

This is uncommon in previous model narratives. The industry previously only looked at benchmark scores; now, latency and throughput are placed on shelves as luxury goods with clear price tags.

Why?

Because for those letting agents write code, a few seconds of lag can completely shatter the hard-won flow state.

Speed has become the most expensive premium.

But OpenAI is a company with severely strained compute resources.

The $200 Pro tier, due to compute congestion, was directly unplugged by officials on September 10, suspending new purchases for 20 days; on the first day of reopening, announced usage caps were halved, cut from 20x that of Plus to 10x. Meanwhile, a new $500 monthly tier appeared prominently, exclusively enjoying the fastest models.

This is an extremely typical compute rationing system, reserving the scarcest resources for wallets that care least about price, and using sufficiently cheap sub-flagships to defend the mass market.

From Astra on September 3, to Sol on the 22nd, to 6.1 Sol on the 29th, three models were released in one month. Each release essentially redraws a cost red line for the server room.

Small model startups suffered collateral damage. Decisions API compresses the smallest Luna to one judgment every 150 ms, specializing in classification and routing. Some view it directly as a precise assassination of TypeSafe's decision model, Jev.

Universal Account

Space, Pages, slides, team tasks, meeting plugins, plus @ChatGPT stationed in Slack and Teams.

Put these puzzle pieces together, and ChatGPT looks nothing like a chatbot anymore.

It resembles a desktop operating system taking shape.

Lined up directly in front are names like Notion, Google Workspace, Slack, and Microsoft Office. The slides specifically emphasize lossless export to PowerPoint and Google Slides.

More interesting are the partners sitting in the VIP seats.

Notion is one of the first 16 partners for 'Sign in with ChatGPT', allowing users to directly deduct ChatGPT subscription quotas within Notion; but on the same day, OpenAI released Pages and Space, which compete head-to-head with Notion on almost every feature point.

Figma, Adobe, and Salesforce entered the big list of the OpenAI software marketplace, where enterprises can use pre-paid contract quotas for OpenAI to purchase their products; but on the same day, ChatGPT is gradually turning image making, layout design, and customer ticket handling—originally their work—into its own native skills.

'Sign in with ChatGPT' is superficially a password-free login plugin, but 骨子里 it is building a universal account for the AI era. 1.2 billion weekly active users don't need to pay separately in each application; quotas flow through OpenAI, and downstream developers settle bills with OpenAI.

Whoever money flows into first pocket becomes the true landlord.

Chassis

In the first half of this year, the word most often used for courage in the AI venture capital circle was 'harness'.
Entrepreneurs and investors liked using it to soothe a spreading sense of insecurity. No matter how strong the model is, it's just an engine; the entire set outside—the conversation orchestration, context compression, tool retrieval, and fault tolerance recovery that truly enable agents to work—is the real moat.

At that time, everyone firmly believed one thing: big tech only had time to pile up compute and couldn't handle these dirty jobs.

This DevDay shattered this assumption. With the Agents API officially adding computer use, OpenAI served up the entire harness running under Codex, ChatGPT, and dot, skin and bone included. Open-source code, provide management, and pull AWS along to integrate into Bedrock.

Big tech not only did it but made it a factory-standard component. The most universal, standardized layer of the agent framework was packaged and taken by the original manufacturer. Consequently, the developer ecosystem's niche also slid into a subtle position.

Codex Cloud allows offline running in the cloud with the computer closed; code reviews can finish initial checks for you at midnight; Security Cloud scans repositories and fixes issues automatically around the clock; the /agents view in the new CLI lets people command several agents simultaneously to start working.

The act of typing on keyboards is being stripped away. Developers are no longer the ones writing code but sit at workstations hitting Enter to approve.

If universal frameworks no longer constitute a moat, where can startup paths lead?

Either dive deep into industry private domains that big tech cannot touch, or go into dead corners that big tech avoids due to compliance concerns.

However, OpenAI clearly didn't intend to leave much room. On the same day as the launch, zero data retention and privacy inference were presented together.

The doors left for middlemen are closing one by one.

Boundaries

The day before the launch opened, many waiting to see GPT-6.1 Astra got empty-handed.

OpenAI voluntarily halted this flagship originally scheduled for October. Saachi Jain, responsible for the safety system, stated that the model had not fully met release standards regarding 'not crossing authorization boundaries and truthfully reporting operational behaviors to humans'.

This was a rare emergency brake.

In the large model race, everyone was accustomed to jumping the gun, grabbing cards, and competing on launch nodes. When almost the entire industry was swept up in FOMO and sprinting, pulling a ready-to-box flagship model off the stage requires considerable resolve.

Stepping back often requires more determination than charging forward.

But this precisely makes Dot on stage seem logical.

When an agent begins to possess an independent cloud computer, browse the web autonomously 24/7, and call tools, the real technical threshold is no longer about brushing two percentage points higher on Benchmarks.

The difficulty lies in delivering trust.

In every product detail of dot, you can see the layered trade-offs made by engineers.

Read-only proactive research isolates the risk of accidental operations; high-sensitivity operations retain manual approval one by one; third-party services have credentials taken over by system proxies, with passwords never exposed to the model in plaintext.

These rules are defined very finely, appearing somewhat cautious.

But this is the prerequisite for whether enterprises dare to truly hand over invoices, contracts, and codebases. Safety is no longer an ethical manifesto hanging on the official website but a solid commercial admission ticket.

When Altman discussed this defense line on stage, his attitude was calm. He said if alignment is treated merely as an engineering problem, deeper propositions will be missed; regarding various jokes from the outside world, he said if people need to project their anxiety onto someone or joke about him to relieve stress, he doesn't mind.

After sprinting for three years, the industry began to realize that what determines how far a model can go is never just the throughput in the compute cluster, but its sense of propriety when interacting with the real world.

All permissions are finely divided: what goes to machines, what stays with humans.

Like that simple rule written at the bottom of the system manual, Dot can draft contracts, troubleshoot code, and connect 4,000 applications for you, but when it comes to changing passwords, it will quietly stop and wait for humans to do it themselves.

This is a mature beginning.