OpenAI DevDay 2026: dots take on work, Sol cuts the cost

OpenAI wants you to hand over a responsibility and get finished work back. Dots, shared workspaces and a cheaper model for longer tasks made that ambition concrete at DevDay on September 29. For businesses, the important question is how much coordination disappears between asking for something and receiving a useful result.

This is a first look at Tuesday's opening keynote, based on OpenAI's recording and official product documentation. The event continues: the published schedule runs technical sessions until 00:30 Swedish time on Wednesday, with the closing session starting at 01:00. Performance claims below come from OpenAI; Hammer has not yet tested these products.

Source: OpenAI DevDay 2026 schedule.

02:28 – Dots take on ongoing responsibilities

Dots are OpenAI's new always-on AI agents, powered by GPT-6 Astra. An agent can use tools and complete several steps toward a goal. Dots have their own cloud computer and can keep working between conversations. You choose the connected apps and specify when the agent can act independently or needs approval.

OpenAI's examples are concrete: following customer requirements, updating material for a board meeting and preparing code changes for an API migration. They show the work the product is designed for. These are vendor examples whose results require testing in your own business.

Business value: a sales lead could delegate tracking a customer's requirements and receive an updated brief before the next meeting. Measure the time spent gathering information, correcting the brief and following up. That is where a working agent needs to earn its keep.

Access: rolling out to Pro and Business Premium in eligible markets. Enterprise requires an administrator to enable the feature. OpenAI's recap also describes an admin-enabled beta for Edu and Healthcare. Access for a specific Swedish account has not been verified here.

Sources: OpenAI on dots and OpenAI's DevDay recap, archived September 29, 2026, at 17:22 UTC.

08:20 – Space and Pages bring people and agents together

ChatGPT Space is a shared home for documents and project material. Pages are editable documents where colleagues and AI can work together. A page can contain text, charts and interactive tools, with updates from connected sources through a configured automation.

That offers a possible answer to an everyday cost: someone has to copy a useful AI response into the right document and keep it current. A shared project page could hold both the changes and the discussion.

Space is offered on Pro, Business, and Enterprise. Create and edit on web or desktop; at launch, mobile can find, read, and share pages. Collaborative slides and spreadsheets are still coming later.

Source: ChatGPT Space features and availability.

19:39 – GPT-6.1 Sol lowers the cost of longer work compared with Astra

GPT-6.1 Sol launches with what OpenAI describes as near-Astra performance for coding, computer use and professional work. That is the vendor's assessment, not our own comparative test.

Standard API prices for prompts with up to 272,000 input tokens are:

• Input: USD 2 per million tokens.
• Cached input: USD 0.10 per million tokens.
• Output: USD 10 per million tokens.
• Cache writes: USD 2.50 per million tokens.

An illustrative request using 100,000 ordinary input tokens and 20,000 billed output tokens costs USD 0.40. Tools, storage and any regional premiums are excluded. Prompts exceeding 272,000 input tokens incur higher rates for the entire request.

The ordinary input and output rates match GPT-6 Sol. The news is therefore the promise of higher capability at that price point, with cheaper cached input. For a developer, that is a good reason to run the same document task or code change through both models and compare cost per accepted result.

GPT-6.1 Sol supports US and EU data residency. Fast mode is unavailable with EU data residency. Tool calling requires the Responses API.

Sources: GPT-6.1 Sol model and pricing and OpenAI API changelog.

20:33 – Ultrafast sells shorter waits

Ultrafast is a more expensive speed tier. OpenAI claims up to eight times faster token generation in Codex and up to six times in the API. Those figures concern generation; external tools, networking and review still affect how long the whole task takes.

GPT-6 Astra Ultrafast is available in the API and in ChatGPT Work and Codex on Pro 500 and Enterprise. GPT-6.1 Sol Ultrafast is still coming soon. The new Pro 500 plan is described as offering 25 times the Plus usage allowance.

API customers face separate request and token limits. Ultrafast supports global processing and US data residency, with no EU regional processing option. That may determine whether the service fits a Nordic client engagement.

Practical takeaway: pay for speed where a person is actively waiting for the next step. For background jobs, cost per completed task often matters more.

Sources: Ultrafast API documentation and OpenAI's recap, archived September 29, 2026, at 17:22 UTC.

29:12 – The Agents API gains computer use

The Agents API lets developers build with OpenAI's managed agent runtime. OpenAI handles sessions, orchestration, context compaction and recovery. Tuesday's addition is computer use in an OpenAI-hosted browser. The API's public beta itself already appears in the September 10 changelog.

That makes something like a website-testing agent easier to build. Your application still needs to handle website permissions, sign-in and result verification. Model usage, tools and hosted environments are billed at their respective rates.

An important distinction for business buyers: the Agents API currently supports only US data residency and does not support Zero Data Retention. The separate Private Intelligence announcement does not automatically change those terms. Its Private Inference preview is expected this fall.

Sources: Agents API features, billing and data terms, computer use and OpenAI's archived DevDay recap.

45:07 – Plugin extensions give your product a new home

Plugin extensions let developers build a sidebar app, editor, or file viewer inside ChatGPT. For a software provider, that could become another route to users: customers can work with the product alongside their conversation.

The documentation is more specific than the broad launch message on stage: web extensions for Free and Go are coming later, and composer mentions are available only in the desktop app.

OpenAI also presented Sign in with ChatGPT, allowing Plus and Pro allowances to be used in selected partner tools. Consumption counts toward the user's plan limits. This is a partner launch, rather than blanket funding for arbitrary API applications.

Sources: Plugin Extensions developer documentation and OpenAI's recap, archived September 29, 2026, at 17:22 UTC.

What is worth trying first?

Hammer's assessment: the greatest potential lies in recurring work that needs both information gathering and updates to a shared document. Choose a real example, such as the brief for a customer meeting. Compare the current effort with a trial in which an agent prepares the material and a person reviews it.

Measure review time and actual usefulness alongside cost. A polished first draft is a start. A brief that remains useful for the next meeting is the more interesting result.