# What OpenAI announced at DevDay

> OpenAI DevDay 2026 introduced Dots, ChatGPT Space, GPT-6.1 Sol, Ultrafast, new agent APIs, and new ways to distribute software. Launches, previews, prices, and limits.

- Published: 2026-09-29
- Last reviewed: 2026-09-29
- Data as of: 2026-09-29
- Article slug: openai-devday-2026
- Author: [Glevd](https://x.com/glevd)
- Reading time: 13 minutes
- Topics: openai, devday, agents, codex, gpt-6-1-sol, pricing
- Data policy: Dated analysis; static benchmark claims are retained.
- Canonical URL: https://benchlm.ai/blog/posts/openai-devday-2026

OpenAI announced Dots, ChatGPT Space, GPT-6.1 Sol, Ultrafast, and a public beta of the Agents API at DevDay on September 29, 2026.

It also introduced ways to use a ChatGPT subscription in partner apps, build fuller plugin interfaces, and buy software through the OpenAI Marketplace.

Some features launch today, others are previews, and several are coming later. That distinction matters if you're deciding what to try after the keynote.

Our read is that OpenAI is putting more of an agent's working life in the cloud: its tools, ongoing tasks, shared documents, and access to other software. Lower token prices help pay for that work. Faster generation helps when a person is waiting for it.

This recap follows the [opening keynote](https://www.youtube.com/watch?v=Fls_onRviPM) and [OpenAI’s official recap](https://openai.com/index/devday-2026-recap/), with product documentation linked where we checked details. Availability below describes what OpenAI announced, with rollout qualifications where the documentation adds them. We haven't independently tested the new releases.

## Dots keep working between conversations

![OpenAI DevDay stage displaying Dots above four colorful agent characters](/images/blog/devday-2026-dots.png)

*OpenAI DevDay keynote. Screenshot supplied by Glevd.*

Dots are always-on agents powered by Astra. OpenAI described giving a Dot a responsibility, such as monitoring bug reports, preparing a budget cycle, or moving an application off a retiring API, and letting it work through the steps. Each Dot has a cloud computer and browser, can write and test code, and uses connected apps and context about its owner. [OpenAI’s Dots introduction](https://openai.com/index/introducing-dots/) confirms the cloud setup, ongoing work, and usage rules.

Onstage, a Dot gathered product feedback, worked on an app redesign through Codex, and joined Slack conversations to take delegated tasks. OpenAI also described its engineers using Dots to investigate bugs and prepare fixes. OpenAI supplied those examples without a measured success rate across customer projects.

People set boundaries for app access, computer use, and actions. OpenAI said Dots work with more than 4,000 apps across ChatGPT and can use plugins already connected to the account. Access still determines which sources a particular Dot can read and which actions it can take.

ChatGPT Space gives that work a shared home. Pages and files sit together, teammates can collaborate, and a Dot can join a page or respond to a comment. Onstage, shared pages included interactive charts, sheets, prototypes, and a launch FAQ assembled from workplace context. Pages could also carry instructions to update themselves from a Slack channel. [Space’s product page](https://chatgpt.com/features/space/) describes pages as documents people and ChatGPT can edit together. Collaborative spreadsheets and slides are coming soon, distinct from the sheets shown in the demo.

| Announcement | What OpenAI announced about access or timing |
| --- | --- |
| Dots | Pro and Business Premium in eligible markets; Enterprise, Edu, and Healthcare beta requires admin enablement and is off by default; rollout is gradual |
| ChatGPT Space and Pages | All Pro, Business, and Enterprise plans on desktop and web; mobile supports finding, reading, and sharing pages, with creation and editing coming later |
| Conversations with your Dot | Do not count toward ChatGPT usage limits; tasks started or managed in Codex or Work count as usual |
| Text messages to a Dot | Coming soon; distinct from messaging inside ChatGPT or Slack |
| Calls with a Dot | Voice interaction demonstrated; external phone access details need confirmation |
| A team of personal Dots | Future capability; one personal Dot is the starting point |
| Collaborative slides | Coming in the next few weeks; team editing, comments, and PowerPoint or Google Slides export |
| Specialist Dots | Enterprise preview of company-configured agents shared across a team |
| Microsoft Agent 365 integration | Work announced to manage specialist Dots through Microsoft's tools |

[Team Tasks](https://learn.chatgpt.com/docs/enterprise/teams) add shared recurring work, such as weekly project updates or a response to a new message. They run in the cloud using the team’s service account and configured connections. Teams can refine the instructions together. OpenAI announced teams and shared tasks for Business and Enterprise.

[@ChatGPT in Slack and Microsoft Teams](https://chatgpt.com/features/chatgpt-in-slack-and-teams/) lets coworkers add context and follow-up requests in a channel, thread, or direct message. An individual ChatGPT license is not required to participate in enabled channels. Organization settings and connected accounts determine access.

[Meetings](https://help.openai.com/en/articles/20001546-the-meetings-plugin-in-chatgpt) is a macOS desktop beta for Pro and Business that saves personalized notes and action items in Space. Enterprise availability is still limited to an alpha. Meeting audio is deleted when notes are ready; incomplete processing can leave it temporarily on the Mac. Slides remain a separate coming-soon feature.

Specialist Dots extend the idea to a company role, with accounting, marketing, and legal work given as examples. Teams can supply goals and context, review output, and share feedback across the company. OpenAI presented them as a preview, so they shouldn't be grouped with the personal Dot rollout as a generally available feature.

For a first trial, ask whether the Dot can complete a bounded responsibility with evidence you can review. Turning a bug report into a reproducible test and a proposed fix is easier to assess than a promise to take care of the engineering backlog.

## Sol costs less, while Ultrafast charges for speed

GPT-6.1 Sol is the new general-purpose model. OpenAI described it as offering near-Astra capability at a fifth of the price, with particular value for coding and repeated agent work. The launch claim doesn't establish how Sol performs on your workload.

[GPT-6.1 Sol’s model page](https://developers.openai.com/api/docs/models/gpt-6.1-sol) lists these Standard API rates as of September 29. They cover prompts up to 272,000 input tokens, before applicable regional premiums and tool charges.

| GPT-6.1 Sol token type | Price per million tokens |
| --- | --- |
| Input | $2.00 |
| Cached input read | $0.10 |
| Cache write | $2.50 |
| Output | $10.00 |

Sol's 95% cached-input discount applies to reading a reusable prefix. Writing the cache costs more than ordinary input, and longer prompts have separate multipliers. It therefore helps most when enough of the same context comes back. It doesn't cut the output bill by 95%. Our [prompt caching breakdown](/blog/posts/prompt-caching) explains why reuse decides the saving.

Ultrafast is a processing tier, rather than a separate model. OpenAI announced it across the API, ChatGPT, and Codex, starting with Astra. OpenAI said GPT-6.1 Sol support is coming later. OpenAI’s recap specifies up to eight times Standard token generation in Codex and up to six times in the API. The [API guide](https://developers.openai.com/api/docs/guides/ultrafast-mode) describes rate limits and regional restrictions. Astra Ultrafast launches in ChatGPT Work and Codex on Pro 500 and Enterprise, as well as the API.

![DevDay slide announcing Ultrafast availability and an up-to-300-tokens-per-second speed claim](/images/blog/devday-2026-ultrafast-keynote.png)

*OpenAI's keynote slide. The speed figure is a vendor claim. Image: OpenAI. Screenshot supplied by Glevd.*

| Processing tier | Keynote speed claim relative to Standard | Keynote price multiplier |
| --- | --- | --- |
| Standard | Baseline | 1× |
| Fast | About 2× | 2× |
| Ultrafast | Up to 8×, reaching up to 300 tokens per second; recap specifies up to 6× in the API | 6× |

Generation speed and total task time are different measurements. Coding agents still wait for tools, tests, network requests, and review. Faster streams can help an interactive session without making the whole job eight times faster. Compare those measurements separately on the [speed pages](/speed).

OpenAI introduced Pro 500 and reopened Pro 200. [Plan documentation](https://learn.chatgpt.com/docs/pricing) lists them at $500 and $200 per month, with Astra Ultrafast on Pro 500. In the keynote, Pro 500 was described as offering 25 times Plus usage and supporting eligible partner-app usage through Sign in with ChatGPT. Pro 200 retains frontier-model access, with Sol positioned for more frequent use.

Those subscription allowances aren't an API token budget you can convert at the rates above. API charges, purchased credits, and included subscription usage follow different billing rules.

## The new APIs give agents tools and quick choices

Agents API brings OpenAI's managed Codex harness to developers in public beta. OpenAI described hosting, memory, multi-agent controls, and computer use as core components. Its [API documentation](https://developers.openai.com/api/docs/guides/agents-api/overview) confirms managed sessions, orchestration, context compaction, recovery, and optional execution environments.

Applications can give an agent tools and a task, follow progress, and send more input to the same session. OpenAI handles more of the loop that otherwise lives in application code. Developers still choose the tools and the environment in which those tools operate. Agent access doesn't remove the need to decide what it may change.

Decisions API targets a smaller operation. OpenAI previewed a Luna-based API that chooses from predefined options in a fraction of a second. Examples included routing a request, classifying an image, and selecting the next action for an agent. It keeps image understanding and language support while restricting the output to the available choices.

That resembles the job we give Jev in our [LLM Selector](/tools/llm-selector): read a description and answer fixed questions before code computes a shortlist. Our [Jev field note](/blog/posts/what-is-jev) explains that setup. Decisions API deserves a comparison on the same routing or classification workload, including error rate and price, before a buyer treats either one as the better option. OpenAI announced limited preview access today, with broader release planned in the coming days. The keynote supplied no exact Decisions API rate card.

OpenAI also previewed Private Intelligence. It described Zero Data Retention with Private Safety Processing and a separate private-inference capability. Those address different parts of the data path: retaining content for safety review and protecting content during inference.

[Private Safety Processing documentation](https://developers.openai.com/api/docs/guides/private-safety-processing) adds a qualification to the stage description. Protected safety records can be stored in customer-controlled storage and reviewed in a hardware-attested runtime without human access. Customer-controlled safety storage qualifies the stage description of retention. Private Inference is a preview coming this fall. Its eligibility and technical scope still need checking before a team relies on it.

AWS was part of the developer offering too. OpenAI described Bedrock managed agents powered by OpenAI, plus AWS access to its frontier models and Codex in ChatGPT Work. The keynote didn't establish identical availability for every model, region, or customer agreement.

OpenAI reported a 45% reduction in Responses API time to first token and more than 30% faster tool calls and workflows, alongside 100-fold growth in usage over the past year and more than 99% reliability. OpenAI supplied those operating figures without the measurement window or a reliability definition, so they shouldn't be treated as a new contractual uptime guarantee.

OpenAI’s [GPT-6.1 Sol system card](https://deploymentsafety.openai.com/gpt-6-1-sol#forecasting-misaligned-behavior-with-deployment-simulation-of-internal-codex-traffic) gives a narrow check on the “near-Astra” claim. Across 49,650 simulated internal Codex tasks, Sol received 28 misalignment flags at severity 3 or higher, against Astra’s 27 and GPT-6 Sol’s 42. OpenAI cautions that internal simulations do not measure external deployment safety rates.

Sol also showed unwanted persistence in 23.5% of warning-respect rollouts, against Astra’s 17.4%. That test omitted system controls intended to prevent circumvention. Our read is that the similar serious-flag counts deserve attention, but they do not establish equal reliability across agent behaviors. Test whether an agent respects a blocked action, alongside whether it completes the task.

## Codex can carry the task across devices

Codex's cloud update lets a task continue when the laptop closes, with a person picking it up on a phone, in a browser, or on desktop. OpenAI said the cloud environment supports the same tools as local Codex, including plugins and computer use. Its recap lists cloud access on Plus, Pro, Business, Healthcare, Education, and Enterprise.

![Codex mobile interface showing Cloud and a connected MacBook as execution options](/images/blog/devday-2026-codex-mobile.png)

*Codex mobile graphic supplied by Glevd. Image: OpenAI.*

Opening remarks mentioned work already shipped since the previous DevDay: Linux support, multiple folders in one project, and Codex on a phone. Those features are part of the product's current direction, but the speech didn't present each as a first announcement that morning.

Codex Security Cloud was a launch-day release. OpenAI described continuous vulnerability discovery and proposed verified fixes, with automatic deduplication, scheduled scans, and a new interface. The recap also lists access to Daybreak Blue models without a separate Daybreak application. [Setup documentation](https://learn.chatgpt.com/docs/security/setup) describes cloud scans of connected GitHub repositories and review of findings and proposed patches. Proposed fixes still go through code review.

[Code Review](https://learn.chatgpt.com/docs/code-review?surface=app) adds a desktop view of pull requests, summaries, comments, checks, and diffs. People can ask Codex about a change and read automatic cloud reviews. GitHub review is generally available; GitLab support is in preview.

OpenAI also emphasized that the Codex harness is open source. That foundation powers Codex, Work, and Dots, and the Agents API gives developers a hosted route to it. Sharing a harness doesn't make every product's permissions, connections, or pricing interchangeable.

| Codex update or demo | What it established in the keynote |
| --- | --- |
| Cloud execution and handoff | Tasks can continue away from a laptop and be resumed across devices |
| Refreshed CLI | Voice steering powered by GPT-Live, an /agents view, prompt editing, session resumption, and worktree improvements |
| Ultrafast coding | A small application was created and changed onstage |
| Browser and computer use | Astra interacted with a game and a mobile app simulator |
| Appshots | An app screenshot supplied context for testing flows across screen sizes |
| Multimodal robot demo | Astra, GPT-Image 2.5, and GPT-Live 1 were combined in a programmable robot demonstration |

Several voice interactions failed during the live demos, and the presenters switched to typed prompts. That's useful context for a release recap. Successful portions showed the intended workflow, while the interruptions show why a demonstration can't establish production reliability.

Onstage, the robot sequence illustrated how existing text, vision, image, and voice models can be combined. It wasn't an announcement of new weights for every model named in the demo. Decisions API, still a preview, was mentioned as a future fit for choosing fast actions from visual inputs.

## Developers have new routes to users and buyers

Sign in with ChatGPT lets people authenticate to partner applications and, where supported, use the AI usage included in their subscription. OpenAI announced 16 launch partners, with expansion planned. It means a developer can offer an eligible user a way to bring their existing allowance into the app, rather than funding that model usage themselves.

Authentication alone doesn't establish that benefit. [OpenAI's guide](https://learn.chatgpt.com/docs/sign-in-with-chatgpt) separates sign-in from permission to use a plan and lists some partners as sign-in only. Developers need to check which integration they're building, and users need to check which permission they're granting.

Plugin extensions add fuller application interfaces inside ChatGPT and Codex. OpenAI described editors, dashboards, and workspaces that sit alongside a conversation. Its examples included a meetings app, Figma design editing with team comments, and Photoshop features from Adobe. OpenAI named dozens of launch partners.

ChatGPT Sites also gains connected plugins, logins, and data. Team members can sign in with their own credentials and bring their own agent, so the same site can show a personalized experience. Personal logins let a shared site use each visitor’s connected data and permissions.

Distribution changes were announced alongside the interfaces. ChatGPT can suggest a relevant plugin during a conversation and let a user connect it there. Plugin Creator helps build a plugin. OpenAI also said it would simplify submissions with review tracking, clearer fixes, human-review requests, and tool updates that don't require restarting the entire submission.

[MCP Events](https://developers.openai.com/plugins/build/mcp-events) lets plugins deliver updates that trigger work while a user is away. Users choose what to watch and how ChatGPT should respond. OpenAI supports webhook delivery from the proposed specification, rather than every delivery mechanism in the draft.

[Shareable profiles](https://help.openai.com/en/articles/20001539-shareable-profiles-in-chatgpt) collect showcased Sites and activity in one place. The recap describes Business and Enterprise access, but the Help Center says Enterprise profiles are coming soon. We use that narrower availability here. Personal profiles have their own sharing settings, and workspace skill sharing stays within the workspace.

OpenAI Marketplace is the new business purchasing route. OpenAI announced more than 30 partners, including CodeRabbit, Notion, and Vercel. Enterprise customers can use part of an existing OpenAI commitment to buy partner products. OpenAI also said its Baseten partnership would make open-source models accessible through the marketplace starting today.

That's a purchasing announcement, rather than a promise that every partner is included free in an existing subscription. Eligible spend, the amount of a commitment that can be used, and partner terms need to be checked in the actual agreement.

OpenAI put ChatGPT's weekly audience at around 1.2 billion people in the keynote. OpenAI’s total audience says little about the number of people a new plugin should expect to reach. For a developer, the distribution claim becomes useful when relevant users discover, connect, and return to the product.

## Research updates show what OpenAI has tried internally

Researchers explained how OpenAI uses models to improve its systems. Researchers said a continuous model-assisted loop helped optimize the computer-use harness and yielded more than a twofold latency improvement that had already shipped. They also described model-assisted safety training and monitoring.

OpenAI said its models could complete more than a third of day-long internal research tasks without intervention as of July, compared with weak performance on tasks of that length in January. It referred to an AI research intern announced a few weeks earlier. Research Intern was an internal-system update, rather than a newly available API.

Researchers also reported assistance on more than 100 longstanding open math problems and work involving antibiotics, historical languages, energy, and robotics. OpenAI did not supply enough detail to assess those projects individually. They belong in a recap as OpenAI's account of its research use, rather than evidence that a customer can reproduce the same result on demand.

Closing announcements included event sessions and attendee hardware. OpenAI announced another worldwide usage-limit reset. The keynote didn't specify the full account eligibility, affected windows, or expiry. [OpenAI's reset guide](https://help.openai.com/en/articles/20001498-how-banked-codex-resets-work) distinguishes a saved banked reset from an automatic reset, so check the offer shown in your account.

Attendees were also offered a limited-edition Chromatic handheld, alongside a Codex Game Studio session for building retro games. OpenAI teased a later session on its open-source work, including an enterprise harness, without establishing release details in the keynote.

For model buyers, the next useful evidence is a repeatable GPT-6.1 Sol evaluation against Astra on the same tasks, with retries, cache reuse, and total spend logged. For agent buyers, it's a completed responsibility with a record of what the agent changed and where a person had to step in.

## Frequently asked questions

### Can I use Dots immediately?

OpenAI announced Dots on September 29, 2026 for Pro and Business Premium, plus an Enterprise beta. Access rolls out gradually in eligible markets. Enterprise administrators must enable the beta, which is off by default. Check your account and workspace settings; the launch date does not guarantee immediate access.

### Does Ultrafast make an agent finish eight times sooner?

OpenAI advertises up to eight times Standard generation speed, with a keynote claim of up to 300 tokens per second. An agent also spends time on tools, network requests, tests, and retries. Tool execution affects total completion time, so generation speed alone cannot establish an eightfold reduction in task duration.

### Does Sign in with ChatGPT include subscription usage everywhere?

No. OpenAI separates account sign-in from permission to use a ChatGPT plan. Some participating apps support both, while others support sign-in alone. Eligible subscribers can authorize plan usage in supported apps, subject to their plan's limits. An app appearing on a partner list does not by itself establish included AI usage.