OpenAI DevDay 2026: the full picture behind the models, cloud agents and plugin push

The [official recap](https://openai.com/index/devday-2026-recap/) counts more than 20 announcements. OpenAI is building three connected layers: intelligence and execution through models, Codex and APIs; distribution through ChatGPT plugins; and persistent work through dots, Space and team tasks. That is an editorial reading of the releases, not a claim that every feature has reached general availability. This post follows our [first-wave report](https://saignals.buzz/news/openai-devday-2026-first-announcements--91452b2c0995) and adds the post-event view, plan limits and technical detail.

GPT-6.1 Sol: better economics, with a benchmark caveat

**GPT-6.1 Sol** is OpenAI's upgrade to GPT-6 Sol for agentic coding, computer use and professional work. In the API, `gpt-6.1-sol` costs **$2 per million input tokens, $0.10 per million cached input tokens and $10 per million output tokens**. Standard input and output are one fifth of GPT-6 Astra's price; cached input is also half the price of GPT-6 Sol's cached input. That makes repeated-context agent loops a concrete cost case to evaluate, not just a model-ranking exercise. The model is available in the API and in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. OpenAI says it is **not yet available in Chat**. [Official model announcement](https://openai.com/index/introducing-gpt-6-1-sol/)

OpenAI reports that Sol 6.1 matches Astra on DeepSWE 1.1 at roughly one fifth of the cost per task, improves on GPT-6 Sol's best result by 6.4 percentage points, and comes within 2.1 points of Astra on an OSWorld 2.0 offline computer-use set at much lower cost. Those are **OpenAI-reported evaluations**, with specific harnesses and reasoning settings; they are reasons to rerun a team's own tests, not a guarantee that every codebase will see the same result. [Benchmarks and methodology](https://openai.com/index/introducing-gpt-6-1-sol/)

**Ultrafast** addresses latency rather than model quality. The DevDay expansion offers up to eight times faster token generation in Codex, reaching 300 tokens per second, and up to six times faster in the API. The API's Ultrafast concept existed in limited preview before DevDay; the notable change is broader **GPT-6 Astra Ultrafast** availability. It is live in the API at initial rate limits and in Work/Codex for Pro 500 and Enterprise. **GPT-6.1 Sol Ultrafast remains upcoming.** The [API guide](https://developers.openai.com/api/docs/guides/ultrafast-mode) says it costs more than standard processing and recommends WebSockets for multi-turn agents to keep network overhead from consuming the speed gain. Do not confuse it with the earlier Fast tier. [DevDay recap](https://openai.com/index/devday-2026-recap/)

Codex Cloud turns a coding session into a team environment

**Codex Cloud** gives each coding task its own remote workspace built from a reusable environment containing repositories, dependencies, tools and approved access. A task can continue while the developer's laptop is asleep, with work reviewed from web, mobile or desktop. OpenAI lists Plus, Pro, Business, Healthcare, Education and Enterprise access. This is the operational change behind the cloud announcement: teams can standardize setup once and send repeated tasks into the same prepared context, while reviewing diffs and tests before committing or opening a pull request. [Codex Cloud documentation](https://learn.chatgpt.com/docs/cloud)

OpenAI also refreshed the **Codex CLI** with voice steering, an `/agents` view, better prompt editing, session resumption and worktree handling. In the desktop app, a new code review flow provides summaries and diffs across projects, while automatic first-pass reviews can run in the cloud. The recap lists the CLI and review features across plans. These tools address different steps: task execution, coordination and review. [DevDay recap](https://openai.com/index/devday-2026-recap/), [code review guide](https://learn.chatgpt.com/docs/code-review)

**Codex Security Cloud** adds repository-wide scans for GitHub projects on demand or by schedule, ongoing checks of new commits, investigation and deduplication of findings, and fixes prepared in the cloud. It includes access to Daybreak Blue models without a separate Daybreak application. OpenAI lists Pro, Business, Enterprise and Edu access on desktop and web. The value proposition is sustained security work across a repository; teams still need to review findings and proposed changes. [DevDay recap](https://openai.com/index/devday-2026-recap/), [security setup guide](https://learn.chatgpt.com/docs/security/setup)

Agents API, Decisions API and AWS: three different building blocks

The **Agents API** launched in public beta earlier in September; DevDay's specific addition is **computer use**, so an agent can interact with software. The platform also exposes multi-agent operation, tool search, tool calling and context compaction. OpenAI runs the harness infrastructure and says there is no separate Agents API fee beyond model and tool usage. This is a route for developers who want Codex-like agent behavior inside their own applications, rather than inside the Codex interface. [Agents API announcement](https://openai.com/index/introducing-the-agents-api/), [computer-use documentation](https://developers.openai.com/api/docs/guides/agents-api/tools/computer-use)

**Decisions API** tackles narrower decisions: a developer defines a finite set of possible answers, supplies text or image context, and uses Luna's output for classification, routing or an agent's next action. It is in **limited preview**, with a broader release planned in the coming days. **Bedrock Managed Agents**, built with AWS, combines OpenAI models and the Codex harness with AWS identity, permissions and infrastructure. AWS labels that offering **Preview**. Neither should be reported as universally available production infrastructure today. [DevDay recap](https://openai.com/index/devday-2026-recap/), [AWS product page](https://aws.amazon.com/bedrock/managed-agents-openai/)

The plugin launch is bigger than tool calling

OpenAI opened more of ChatGPT's own interface to plugin developers. **Plugin extensions** can create a sidebar app, a panel beside a conversation, product settings, file viewers and editors, deep links, rich forms and context sharing between the model and app. That lets a developer ship a place to work, not only a tool whose result appears as text. The detailed documentation says web support for Free and Go users is still coming, and composer mentions are currently desktop-only, even though the recap describes extensions as available across plans. [Plugin extensions documentation](https://developers.openai.com/plugins/build/extensions)

The packaging distinction matters. A **plugin** is what users discover, install and share; it can contain a skill, an MCP server, or both. A **skill** describes a repeatable workflow. An **MCP server** supplies live data, authenticated tools and optional UI. OpenAI also announced Plugin Creator, a clearer submission process and improved ranking and recommendations, while users retain control of plugin permissions. For builders, the combination covers development, distribution and the runtime interface. [Plugin architecture](https://developers.openai.com/plugins/concepts/plugins), [DevDay recap](https://openai.com/index/devday-2026-recap/)

**Sites** can host supported plugins for workspaces on Business, Enterprise, Healthcare and Edu plans, with each teammate using their own connected data and permissions. **MCP Events** lets a plugin advertise updates such as new messages or document comments so ChatGPT can start a user-specified automation. The [technical guide](https://developers.openai.com/plugins/build/mcp-events) currently describes webhook delivery, requires MCP 2.0 and follows a **draft** Events specification; polling and streaming are not supported by this integration. That is meaningful for event-driven products, but builders should validate protocol and rollout support before committing to it. [DevDay recap](https://openai.com/index/devday-2026-recap/)

Dots and Space show where persistent agents will work

**Dots** are OpenAI's always-on personal agents. The detailed [dots announcement](https://openai.com/index/introducing-dots/) says they use GPT-6 Astra, have their own cloud computer, connect to apps, can work across ChatGPT, Slack and Teams, and can be inspected and redirected. Proactive background research uses read-only connected tools; actions can require review or approval. The first dot is rolling out to Pro and Business Premium in eligible markets. Enterprise, Edu and Healthcare workspaces can try a beta when an admin enables it. Specialist dots with separate identities are in focused enterprise pilots. These plan limits matter more than the broad 'always-on' slogan.

**ChatGPT Space** is the shared setting for teammates, ChatGPT and dots. **Pages** are collaborative documents where people and agents can work together. **Team tasks** can repeat on a schedule or respond to events from connected tools; OpenAI lists them for Business and Enterprise. @ChatGPT in Slack and Microsoft Teams brings connected tools into team conversations. A Meetings plugin is in macOS beta for Pro and Business, with Enterprise later. Collaborative slides are **promised for the coming weeks**, not available on launch day. [DevDay recap](https://openai.com/index/devday-2026-recap/), [Space overview](https://chatgpt.com/features/space/), [teams documentation](https://learn.chatgpt.com/docs/enterprise/teams)

Privacy, plan usage and commercial access

OpenAI's **Private Intelligence** announcement includes Zero Data Retention with Private Safety Processing, which supports automated safety review without retaining customer prompts or responses in OpenAI systems. The [technical guide](https://developers.openai.com/api/docs/guides/private-safety-processing) describes customer-managed storage and setup responsibilities. **Private Inference**, based on confidential computing, is a preview planned for this fall. Treating the two as a single live product would overstate today's availability. [DevDay recap](https://openai.com/index/devday-2026-recap/)

The access model is changing too. **Sign in with ChatGPT** provides identity and, in participating tools, lets eligible Plus and Pro users draw on plan allowance; OpenAI names 16 launch partners. **Pro 500** offers 25 times the Plus allowance and Ultrafast access. For enterprises, the **OpenAI Marketplace** lets eligible customers express interest in applying existing commitments to approved partner software; the first group has 32 partners. These are commercial and distribution changes, not model capabilities, and teams should check their own plan's eligibility. [DevDay recap](https://openai.com/index/devday-2026-recap/), [plan-usage guidance](https://help.openai.com/en/articles/20001542-using-your-chatgpt-plan-in-other-apps-and-sites)

What is live, limited or still promised?

| Available now, subject to plan or rate limits | Preview or limited access | Announced for later |
|---|---|---|
| GPT-6.1 Sol in API, Work and Codex; Astra Ultrafast; Codex Cloud; Codex Security Cloud; plugin extensions with surface restrictions; Space and Pages | Decisions API limited preview; Bedrock Managed Agents preview; enterprise dots beta and specialist-dot pilots; Meetings plugin macOS beta | GPT-6.1 Sol Ultrafast; Private Inference preview this fall; collaborative slides in coming weeks; broader rollout of some plugin surfaces and dots |

This grouping follows the [official recap](https://openai.com/index/devday-2026-recap/) and linked product documentation as of September 30. It does not imply that an individual account has access to every item in its plan; regional, admin, beta and rate-limit controls still apply.

What developers should do next

The practical sequence is to **measure before migrating**. Benchmark Sol 6.1 on real agent tasks and compare cost per completed task, not only token rates. Pilot Codex Cloud with a reusable environment and review its permissions and outputs. Treat computer use as an application capability requiring clear boundaries and evaluation. If shipping a plugin, start with the smallest combination of skill, MCP tools and UI that solves the workflow, then test each supported ChatGPT surface and any event trigger. Finally, label preview features in product plans. DevDay provided the components for a broader agent platform; the quality of the resulting systems will depend on their implementation, controls and measured performance.