Everything OpenAI announced at the DevDay 2026 keynote

OpenAI's DevDay 2026 keynote put always-on agents at the center of ChatGPT, introduced GPT-6.1 Sol, and expanded the tools developers can use to build, run and sell AI applications. The announcements on September 29 at San Francisco's Fort Mason covered much more than a model release: Dots, a shared workspace called Space, faster inference, cloud coding, browser automation and new ways to bring existing ChatGPT subscriptions into other products.

The connecting idea was continuity. OpenAI wants an agent to keep working after a conversation ends, carry its context between devices, and deliver its work into the documents and applications a team already uses. Cheaper models, faster responses and reusable cloud environments are the infrastructure behind that ambition.

This recap draws on our keynote transcript, our live coverage with announcement posts and presentation clips, and the product documentation published alongside the event. Availability below reflects the launch-day documentation; a rollout announcement does not mean every eligible account already has access.

The biggest announcements, at a glance

• Dots: personal agents that remain active between conversations, with their own cloud computer and access to approved connected tools.
• ChatGPT Space: shared pages and files where people and agents can work together. Collaborative slides and spreadsheets are still coming.
• GPT-6.1 Sol: a more capable everyday model for coding and professional work, with standard API prices of $2 per million input tokens and $10 per million output tokens.
• Ultrafast and Pro 500: a premium speed option for Astra, and a new $500 monthly subscription. Sol support for Ultrafast is coming later.
• Codex and agent infrastructure: reusable cloud environments, expanded CLI controls, security scanning, hosted browser use and an AWS managed-agent preview.
• Distribution: Sign in with ChatGPT plan usage, app-like plugin extensions, event-driven tasks, and an enterprise Marketplace.

Jump to: Dots · Space and collaboration · Sol and pricing · Ultrafast and Pro · Codex · Agent APIs · Privacy · Plugins and distribution · The demos · Quick answers

Dots: an agent that stays on the job

Sam Altman opened the product announcements with Dots, describing a shift from asking an assistant isolated questions to giving it an ongoing responsibility. His examples included monitoring bug reports, preparing a budget cycle and migrating an application before a legacy API shuts down.

A Dot has its own cloud computer and browser. It can use connected plugins, research a problem, write and test code, and bring work back for review. OpenAI said Dots can work across more than 4,000 connected applications. The intended handoff is closer to assigning a project than prescribing every click: investigate the dependencies, work through the changes, and return a pull request or another result a person can inspect.

OpenAI's launch documentation says Dots are powered by GPT-6 Astra. The initial rollout covers Pro and Business Premium in eligible markets. Enterprise, Edu and Healthcare access is a beta that administrators must enable; it is off by default. Specialist Dots for companies are being piloted separately.

Permissions matter to how this works. People choose the connected tools and boundaries a Dot can use, and consequential actions require review. OpenAI describes proactive background research as using read-only tools. Talking to a Dot does not itself consume the Work and Codex allowance, but delegated tasks can. A Dot included with a subscription is therefore not a promise of unlimited background computation.

The keynote also previewed company-configured specialist Dots with their own identities and credentials, and work with Microsoft on Agent 365 integration. Text messaging and a broader future in which users manage multiple Dots were presented as next steps, rather than features everyone receives immediately.

Follow the Dots announcement in our live timeline ↗

Space gives the work a shared home

The other large ChatGPT launch was Space, which replaces Library for eligible users and brings pages and files into a shared working area. Pages can contain writing, research, visualizations and code. Teammates can edit together, leave comments and mention ChatGPT or a Dot to request changes.

The useful distinction is that a page can remain current. Instead of repeatedly asking for a new summary and pasting it into a document, a team can ask an agent to monitor an approved source and keep the page updated. That makes Space a destination for ongoing work, rather than just a storage folder for finished outputs.

At launch, Space is available on Pro, Business and Enterprise. Creation, editing and collaboration are on web and desktop; mobile supports finding, reading and sharing pages. Mobile editing, collaborative slides and collaborative spreadsheets are coming later. Existing spreadsheet or presentation files in Space should not be confused with those future collaborative editors.

See this announcement in our live coverage ↗

OpenAI also expanded the surrounding collaboration tools. Team tasks support shared scheduled or triggered work under workspace controls. ChatGPT in Slack and Microsoft Teams lets coworkers bring the assistant into an existing discussion, subject to the organization's setup and permissions.

The Meetings plugin turns meeting audio into notes and suggested follow-ups without a bot joining the call. Its initial beta is in the macOS desktop app for Pro and Business; Enterprise has a limited alpha, and other operating systems are coming. Notes are private by default, and the recorded audio is deleted after processing rather than retained for playback.

Finally, shareable profiles give users a place to display selected Sites and their work. Personal profiles start private. The web and desktop showcase does not expose private conversations, and Enterprise support is still coming.

GPT-6.1 Sol: the cost-and-capability announcement

GPT-6.1 Sol was the keynote's main new general-purpose model. Altman's pitch was near-Astra capability at a lower cost, particularly for the repeated coding and tool-use steps that accumulate inside a long agent task.

The standard API rates are:

• Input: $2 per million tokens.
• Cached input: $0.10 per million tokens.
• Output: $10 per million tokens.

Those are standard processing rates, not a universal bill for every service tier. Long-context requests and premium processing can change the price. The especially low cached-input rate matters when an agent repeatedly reuses the same project context; cache writes and newly added content have their own costs.

See this announcement in our live coverage ↗

In its model announcement, OpenAI reported 75.2% on DeepSWE v1.1 at high reasoning effort, compared with GPT-6 Sol's best reported 68.8%. It also reported roughly a 32% reduction in responses containing factual errors on deliberately difficult factuality prompts. These are company evaluations under particular conditions, not guarantees that every application will see the same improvement.

Sol is available through the API as gpt-6.1-sol, and in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu. OpenAI explicitly says it is not in regular Chat at launch. That distinction is easy to miss in the broader claim that the model is available on paid plans.

This was a Sol release, not the launch of GPT-6.1 Astra. The keynote used the existing Astra model throughout several demonstrations. Our earlier reporting on the separate Astra release decision remains a different story.

Read the Sol benchmark update in our live timeline ↗

Ultrafast trades a higher price for less waiting

Ultrafast is a speed tier, not a separate intelligence model. OpenAI demonstrated it with Astra, using matching prompts to build and launch a small rocket scene. The faster version completed visibly earlier in the stage comparison.

OpenAI advertises up to eight times faster token generation in Codex, reaching 300 tokens per second, and up to six times faster in the API. That measures generation speed; it does not mean an entire job with browsing, compilation and tests finishes eight times faster. The API guide also describes premium pricing and launch rate limits. Astra is supported now; GPT-6.1 Sol support is coming.

See this announcement in our live coverage ↗

The subscription counterpart is Pro 500, at $500 per month. The keynote described 25 times the usage of Plus and access to Ultrafast. OpenAI also reopened Pro 200 sign-ups. Its plan documentation lists Pro 100 at $100, Pro 200 at $200 and Pro 500 at $500 monthly; among those plans, only Pro 500 includes Ultrafast.

There is a material qualification for Pro 200: new subscriptions without grandfathering receive a lower included allowance than the earlier plan. Eligible existing subscribers retain their previous allowance through October 29, 2026, before moving to the updated allowance. The $200 price stays the same. Readers comparing plans should compare usage and access, rather than assuming the reopened tier has its old terms.

Codex moves more of the work into the cloud

Codex Cloud lets a task continue without keeping a laptop open. Developers can configure reusable environments with repositories, dependencies and approved credentials, then start or continue work from the app, web or mobile. Each task gets its own workspace. The goal is to make the cloud environment useful for substantial projects, instead of repeatedly rebuilding the setup for a single short task.

In his demonstration, Romain moved between local work, remote access and a cloud task, including a request to rewrite an application's backend in Rust. That was the start of a longer job, not an onstage proof that an entire backend rewrite had completed. The practical benefit is persistence and handoff between devices. See the cloud documentation and our separate Codex Cloud report for setup details.

See this announcement in our live coverage ↗

The Codex CLI gained more controls for parallel work: /agents to inspect and switch between running tasks, /fork to branch work with its context, and /usage analytics. OpenAI's announcement also listed themes, managed worktrees, collapsible output, richer diffs, and Mermaid and LaTeX rendering. Voice interaction was part of the demo, although a CLI voice connection failed and Romain continued with a typed prompt.

Code Review brings pull-request inspection into the desktop workflow, including files, checks, comments and review drafting. GitHub support is generally available; GitLab is in preview. The important handoff remains the review: generating suggested feedback does not mean a person has accepted or posted it.

Codex Security Cloud adds cloud-based vulnerability scanning with scheduled and continuous checks, deduplication, and proposed fixes for review. OpenAI also announced access to cyber-capable Daybreak Blue models through the service. Repository permissions and setup still apply; a verified proposed patch is not the same as automatically deploying a fix. Security scanning belongs alongside the team's existing review and release process.

See this announcement in our live coverage ↗

Hosted browser use, fast decisions and an AWS option

OpenAI presented the Agents API as a way to use more of the infrastructure behind its own agents: the harness, hosting, memory and multi-agent controls. Its release history places the initial public beta earlier in September; the DevDay update adds computer use rather than marking the first appearance of the API.

With hosted computer use, an agent can open a browser, inspect pages and work through a website flow. The application remains responsible for handling access approvals and sign-in. A useful immediate application is testing a web workflow without building an entire browser-execution environment yourself.

The new Decisions API tackles a narrower problem: choosing quickly from a predefined set of answers. The keynote described routing requests, classifying images and selecting an agent's next action. It uses Luna and can accept visual inputs. OpenAI's event recap describes a limited preview, with broader access coming. This is not a replacement for a model writing an open-ended report; its value is a small, fast decision inside a larger system.

See this announcement in our live coverage ↗

For organizations already on AWS, Amazon Bedrock Managed Agents powered by OpenAI provides another deployment path. AWS labels it a limited preview and describes agents running within the customer's AWS environment, with inference on Bedrock. It combines OpenAI's harness with AWS infrastructure and existing services. The preview label matters: the keynote partnership announcement should not be read as unrestricted general availability.

See the AWS managed-agent announcement and keynote frame ↗

Altman also reported broader API infrastructure progress: 100-fold growth in Responses API usage over the preceding year, a 45% reduction in time to first token, and more than a 30% improvement in tool calls and workflows. Those figures were OpenAI's own operational claims. They help explain why the company is packaging more long-running execution into managed services, but they are not a service-level guarantee for an individual application.

Private Intelligence has two different timelines

OpenAI previewed Private Intelligence as a way to give enterprise customers stronger privacy controls while retaining model-safety checks. The announcement includes zero data retention with Private Safety Processing, and a separate Private Inference preview planned for later in the fall.

The safety-processing design does not mean no data exists anywhere. Its documentation describes encrypted safety records stored in customer-controlled cloud storage, with a 30-day retention window and an attested automated review process. OpenAI does not retain the customer's prompt and response for that process, and human access is disabled. That is a specific architecture for approved customers, not a blanket promise that every API request now operates under the same arrangement.

For buyers, the distinction is concrete: identify who holds the records, who can decrypt them, and which product is actually available. Private Safety Processing and the future Private Inference offering should be evaluated separately.

Read our original Private Intelligence keynote update ↗

A ChatGPT subscription can travel into other products

Sign in with ChatGPT now does more than identify a user in participating products. Eligible users can opt to put supported AI requests against the Work and Codex usage included with their Plus or Pro plan. Altman named 16 launch partners for this plan-usage announcement and said OpenAI would expand it.

The usage documentation separates signing in from granting access to a plan. Users can set a weekly limit per app; using credits after included usage runs out requires a separate opt-in. The third-party product can still charge for its own subscription or services. This is not unlimited AI usage, and it does not hand an app access to a user's private ChatGPT conversations.

Follow the Sign in with ChatGPT plan-usage update ↗

For developers, the potential benefit is a lower barrier to trying a product: a customer who already pays for eligible AI usage can bring it along. The developer still needs to explain what its application does and why it deserves the user's time and permissions.

Plugin extensions turn tools into fuller applications

Plugin extensions let developers build interfaces in ChatGPT and Codex, including sidebar applications, conversation panels and custom file viewers. The keynote showed Figma design work and Photoshop features inside ChatGPT, illustrating a move beyond a tool that simply returns text.

Developers also get improved discovery in the context of a conversation, a simpler submission process, and Plugin Creator tooling. Sites can use connected plugins and data, allowing the same shared application to behave differently for users with different approved connections. Our plugin report covers the changes in more detail.

MCP Events supplies the event-driven piece: a connected service can notify an agent when something relevant happens, instead of requiring it to repeatedly poll. The user chooses what to watch and what the agent should do. This is the plumbing that can turn a new ticket, document change or application event into a useful follow-up task.

Read the plugin extensions and event-triggered automation update ↗

Marketplace connects developers with enterprise spending

The OpenAI Marketplace gives enterprise customers a way to use part of an existing OpenAI commitment for eligible partner products. The keynote named CodeRabbit, Notion and Vercel among more than 30 partners; OpenAI's recap specifies 32 at launch.

A Baseten partnership also brings open models into that purchasing channel. That is an access and distribution arrangement, rather than a claim that OpenAI released a new open-weight model during the keynote. See our Baseten coverage for the commercial details.

What the demonstrations showed—and what they did not

Holly, introduced as a member of OpenAI's product team, demonstrated Dottie using a fictional music application called Blossom. The scenario connected calendar changes, test-user feedback, a last-minute design revision, a shared page and an engineering bug report. It showed the intended workflow across applications; Blossom was a demonstration setting, not a new music product being launched.

Romain's segment combined a venue-inspired 3D scene, an Astra Adventures game, app inspection in a simulator and a long cloud coding task. It ended with Lavender, a programmable robot loaned by Hugging Face, combining Astra vision, GPT Image 2.5 and GPT Live 1. The robot was an illustration of combining existing capabilities through APIs, not an announcement of a new OpenAI robot or the first release of those image and voice models.

The live presentation also exposed some friction. Holly's voice exchange took longer than expected, and Romain's CLI voice connection failed before he switched to typing. Those moments do not erase the capabilities, but they are a useful reminder that a stage demonstration is not an independent reliability study. Performance claims, access restrictions and the quality of a finished task still need to be checked in the workflow where someone will actually use it.

Research lead Tejal described OpenAI using models to improve its computer-use harness and post-training work, including faster execution and safer behavior in its internal tests. The significance was the feedback loop: agents helping researchers improve the systems that power subsequent agents. The keynote did not establish that research itself has become fully autonomous.

Read the research segment in our live coverage ↗

What changes after this keynote

The most consequential change is the combination. A faster or cheaper model alone improves a conversation. A cheaper model inside a persistent cloud agent, with event triggers and a shared place to deliver results, changes how a team can hand off work.

For an individual, that might mean delegating a recurring research or coding responsibility. For a team, it could mean keeping a shared project page current or turning a reported bug into a proposed patch. For a developer, it means more managed infrastructure and more distribution inside a platform Altman said now reaches roughly 1.2 billion weekly users.

The unanswered question is how consistently these systems finish useful work with the permissions and review steps a real organization requires. Cost per token helps, but cost per accepted result matters more: retries, external tools, human review and failed runs all belong in that calculation. Space and Dots make the result easier to inspect; they do not remove the need to inspect it.

OpenAI closed with additional sessions, a global banked usage reset and a message about giving people more capacity to create. Our DevDay live page preserves the announcement timeline, original posts and captured presentation moments. This article is the organized keynote recap; that timeline remains the place to follow subsequent updates and discussion.

Quick answers about DevDay 2026

What did OpenAI announce at DevDay 2026?

The headline launches were Dots, ChatGPT Space, GPT-6.1 Sol, Ultrafast and Pro 500, expanded Codex Cloud and security tools, agent API updates, plugin extensions, ChatGPT plan usage in partner products, and the OpenAI Marketplace. Several supporting features remain previews or are coming later.

Can I use GPT-6.1 Sol in regular ChatGPT chat?

Not at launch. OpenAI lists it in ChatGPT Work and Codex on eligible paid plans, and through the API. Its standard API rates are $2 input, $0.10 cached input and $10 output per million tokens.

Are Dots and every Space feature available to everyone?

No. Dots have plan, market and administrator restrictions. Space launches on Pro, Business and Enterprise, with mobile editing and collaborative slides and spreadsheets still coming. Check the linked product documentation for your account's access.

Where can I see the original announcements and keynote clips?

Visit Superpower Daily's OpenAI DevDay live coverage for the timestamped timeline, embedded announcement posts, presentation excerpts and discussion. OpenAI's official recap links to the launch pages and developer documentation.

Reporting note: We reviewed 78 published live updates and the captured keynote transcript, then checked launch details against official product pages, API documentation and help-center articles. Company benchmarks and operational figures are attributed to OpenAI. Where headline announcements and detailed documentation differ in specificity, this recap uses the more specific launch-day access rules.