OpenAI said DevDay 2026 contained more than 20 announcements. That is technically true, but it is a terrible way to decide what matters.
The useful version is simpler. OpenAI has made ChatGPT a place where agents can keep working, given teams a shared document-and-task layer, lowered the price of near-frontier model work, and made its coding agent more portable. Then it attached a premium speed lane for customers who value time over token cost.
That is not one product launch. It is a new operating model: ChatGPT as the surface, agents as the workers, Spaces as shared context, and metered compute as the limiting resource.
This guide separates the durable launches from keynote theatre. It also separates a published price from a plan that merely includes usage, because those are very different promises.
The short version: what actually changed
| Launch | What it is | Availability at publication | Published price or billing rule |
|---|---|---|---|
| Dots | Always-on ChatGPT agents with a cloud computer and connected apps | Gradual rollout to Pro outside the EEA, Switzerland and UK; Business Premium; Enterprise beta is admin-controlled | Included temporarily for eligible plans; OpenAI says usage terms after the first month are still to come |
| ChatGPT Space and Pages | Shared files, editable pages and collaboration context | Rolling out; workspace controls apply | No separate price published |
| Teams and team tasks | Shared ownership of scheduled or event-triggered agent work | Business and Enterprise | Uses plan allowance and/or workspace credits, depending on contract |
| Meetings plugin | Mac app meeting notes that flow into Space | Beta for Pro and Business; limited Enterprise alpha | No separate price published |
| GPT-6.1 Sol | Near-Astra model for coding, computer use and professional work | API, ChatGPT Work and Codex; plan/admin availability varies | $2 input and $10 output per 1M standard API tokens |
| Astra Ultrafast | Highest-speed processing lane, initially for GPT-6 Astra | Pro $500 and eligible Enterprise/Edu in Work and Codex; availability varies | 6× Astra Standard on credits; API list price is $60 input and $300 output per 1M tokens |
| Codex Cloud | Isolated OpenAI-managed environments for coding tasks | Plan and workspace controls apply | No standalone price; consumes supported Codex/Work usage |
| Plugins and Sites | Build or host MCP-backed tools, then use or share them in ChatGPT/Codex | Hosting is available to all plans; sharing is policy-controlled in workspaces | No separate platform price published |
| Marketplace | Enterprise program for qualifying partner products | Initial managed rollout | No self-service checkout or public catalogue price |
The table is deliberately less exciting than a 20-item keynote list. It is more useful. Availability is not universal, and several products do not yet have a price card.
Dots are the headline—and the largest trust decision
Dots are OpenAI’s new always-on agents inside ChatGPT. A dot uses GPT-6 Astra, has its own cloud computer, can work with apps a user connects, and can retain context for ongoing work. The important distinction is that it can make progress between conversations: it can run scheduled work, research in the background and surface results for review.
That makes a dot materially different from a chat thread, but not magically autonomous. Users set goals and rules; the agent should bring decisions needing human judgment back to the owner. OpenAI’s own guidance says a dot can make mistakes and that important work still needs review.
At launch, dots are rolling out to Pro users outside the EEA, Switzerland and the UK, Business Premium users in supported regions, and eligible Enterprise workspaces where an administrator enables the beta. Enterprise access is off by default. Texting is a separate, limited US Pro beta, and local-computer access remains opt-in.
The launch price is unusually incomplete. Eligible Pro, Business and Enterprise users get dots without plan allowance drawdown for the next month; OpenAI has not published what the ongoing consumption rule will be. Treat “included” as a limited launch condition, not an unlimited-use guarantee.
The privacy point deserves equal weight. Disconnecting an app does not erase information a dot already obtained; deleting the dot is required to remove its saved memories. A team should decide what an agent may read, write, buy or share before connecting email, Slack or a local machine.
Space, Pages, Teams and Meetings turn ChatGPT into a work area
ChatGPT Space replaces Library as the home for files and Pages. Pages are editable, shareable documents that can hold charts, images, checklists and dashboards; people collaborating on one can each use their own agent against that shared material. Sharing a Page does not grant access to a person’s private chats or Memory, but anything an agent writes onto a shared Page becomes visible to the people permitted to read it.
That sounds like a small interface change. It is actually the connective tissue for the rest of the launch. Dots need a place to leave work, Teams need a place to share context, and meeting notes need a place to become follow-up tasks.
For Business and Enterprise, Teams add a shared owner for Pages, Spaces and recurring work. Team tasks can be scheduled or started by supported events in connected apps, then run in the cloud using approved tools and a team’s configured connections. This is useful for routine briefs, monitoring and draft updates, but it also means the team’s service account, permissions and task instructions matter more than any one person’s chat history.
Meetings is the clearest early example. It is a macOS desktop-app plugin that records a call or in-person conversation, produces a personalized summary and saves suggested action items into Space. It is in beta for Pro and Business, with an Enterprise alpha; OpenAI says iOS, Android and Windows support is coming later. It does not put a bot into the meeting. It does, however, stream microphone and system audio to OpenAI for processing, so participants still need to be told and consent where required.
None of these collaboration features has a separately posted per-user surcharge. That is not the same as free. For credit-based workspaces, the agentic work that follows from them is billed through the applicable plan allowance or credit pool.
GPT-6.1 Sol is the clearest developer launch
GPT-6.1 Sol is OpenAI’s new balance point: the company positions it near GPT-6 Astra on complex coding, computer use and professional tasks at one-fifth of Astra’s Standard input and output prices. It has a 1.05 million-token context window, a 128,000-token output ceiling, image input, and support for tools including web search, file search, image generation, hosted shell, computer use, MCP and skills through the Responses API.
It is not a drop-in “turn reasoning off” model. Sol supports low, medium, high, xhigh and max effort; none and minimal are unavailable. Tool calling requires the Responses API, while Chat Completions support is limited to requests without tools. It supports US and EU data residency, although Fast mode is not available with EU residency.
Here is the pricing that should drive an API buyer’s test plan. These are dollar prices per million tokens for prompts up to 272,000 input tokens.
| GPT-6.1 Sol API processing | Input | Cached input | Cache write | Output |
|---|---|---|---|---|
| Standard | $2.00 | $0.10 | $2.50 | $10.00 |
| Batch or Flex | $1.00 | $0.05 | $1.25 | $5.00 |
| Fast | $4.00 | $0.20 | $5.00 | $20.00 |
Longer prompts cost more: OpenAI says a request above 272,000 input tokens is priced at twice the input and cache rates and 1.5 times the output rate for the entire request. Regional processing adds 10% where it is offered. That means the headline $2/$10 rate is real, but context design still determines the bill.
For credit-based ChatGPT Business and Enterprise/Edu workspaces, Standard GPT-6.1 Sol is listed at 50 credits per million input tokens, 2.5 cached-input credits and 250 output credits. OpenAI says a typical Codex task on Sol can consume roughly two to 15 credits. Those are planning ranges, not a cap.
Ultrafast is a speed product, not a cheaper model
OpenAI also introduced Ultrafast processing for GPT-6 Astra. The useful claim is not that every task should run there; it is that latency-sensitive development work can buy a much quicker feedback loop.
The cost is blunt. API list pricing for short-context Astra Ultrafast is $60 per million input tokens, $6 cached-input tokens, $75 cache writes and $300 output tokens. That is six times the Standard Astra price. For credit-based Work and Codex usage, OpenAI likewise lists Ultrafast at 6× the Standard Astra rate and says it consumes included allowance at 8× the rate.
At publication, official plan guidance identifies Ultrafast in ChatGPT Work and Codex for the $500 Pro tier and eligible Enterprise and Edu plans. OpenAI has not published a general consumer price breakdown beyond identifying that plan. Sol Ultrafast is described as coming later, so do not assume a $500 subscription turns every model into the fast lane.
The buyer question is simple: does a shorter wait improve a valuable enough human workflow to justify a 6× model charge? For interactive coding and incident work, perhaps. For overnight research, batch jobs and document processing, almost certainly not.
Codex Cloud, plugins and Marketplace widen the platform
Codex Cloud moves coding tasks onto OpenAI-managed computers. A prepared environment contains a project’s repositories, tools, dependencies and access settings; each task runs in an isolated workspace. A developer can start or continue a task from desktop, web or mobile, and cloud work can keep running while their own computer sleeps.
That is real convenience, but it is not a substitute for governance. OpenAI notes that Codex Cloud is not covered by its BAA and should not process protected health information. Environments can contain shared resources and credentials, so teams need to inspect what an environment exposes before sharing it. Existing Code Review, Security Review and some integrations remain in the legacy cloud experience during the transition.
The developer-platform update is bigger than Cloud. A builder can now host an MCP server in a new or existing ChatGPT Site, have ChatGPT or Codex create a plugin around those tools, then install or share it under workspace controls. That makes a Site more than a document surface: it can become a controlled interface to a handbook, dashboard or other internal system.
Marketplace is the enterprise-commercial complement. It lets eligible enterprise customers apply part of their OpenAI commitment to qualifying partner products built on OpenAI models. In the initial rollout, this is not an app store checkout: OpenAI and the partner determine eligibility, the partner contracts and invoices the customer, and OpenAI reconciles the allowed amount against the commitment. There is no public product price list to compare yet.
OpenAI also moved workspace controls into an Agent Security area and added local Work sync. The latter lets a person continue a new local Work task from web or mobile while the connected computer remains online. It is a useful continuity feature, but synced task context is coordinated in the cloud and zero-data-retention workspaces are not supported.
The announcement that needs a caveat
Independent DevDay coverage also describes a new Decisions API for narrow, repetitive business choices. That could become a meaningful developer product, but OpenAI had not published an official API reference, availability terms or price at the time of writing. It belongs on the DevDay announcement list, not in a procurement spreadsheet yet.
The same restraint applies to the withheld GPT-6.1 Astra release. Reporting around DevDay says OpenAI paused that more advanced model over security concerns. A paused model is consequential news, but it is not a product you can deploy today.
What the day means
DevDay’s most important change is not a single model score or a charming agent name. It is the conversion of ChatGPT from a destination for prompts into a place where work can persist, be shared and consume metered compute.
For an individual, the sensible first experiment is a bounded dot task or a small Codex Cloud environment with permissions you understand. For a team, start with a Space, a narrowly scoped Team task and a clear review owner before connecting systems that can send messages or touch files. For an API buyer, benchmark GPT-6.1 Sol against the present workflow with realistic cache reuse and output lengths before paying for Ultrafast.
OpenAI shipped a lot. The value will come from knowing which of those launches is a real capability for your account, what data it can see, and what every extra minute of agent work costs.
Support the site at buymeacoffee.com/gearpulse.site.