The Short Answer
At DevDay 2026 on September 29, OpenAI launched more than twenty products. The four that matter for a business are dots (always-on agents with their own cloud computer), GPT-6.1 Sol (near-Astra performance at one-fifth of Astra's price), the Agents API with computer use, and a set of workspace features that turn ChatGPT into the place a team works rather than a tool it opens.
Dots got the keynote. Sol changes the budget. And getting value from an always-on agent depends less on the features on stage than on which work you hand it, and in what order.
Introducing dots, powered by GPT-6 Astra. Remarkably capable, always-on agents built to handle everything.
At a Glance
| Launch | What it is | Who gets it |
|---|---|---|
| Dots | Always-on agent on GPT-6 Astra, own computer and browser, 4,000+ app connections | Pro and Business Premium now; Enterprise beta, off by default |
| GPT-6.1 Sol | $2 in / $10 out per million tokens, $0.10 cached | API, ChatGPT Work and Codex, today |
| Agents API | Managed agent runtime with computer use and subagents | API; also Bedrock Managed Agents on AWS |
| Space, Pages, slides, team tasks | Shared workspace where people and agents edit together | Pro, Business, Enterprise |
| @ChatGPT in Slack and Teams | Mention it in a channel; teammates need no license | Business, Enterprise |
| OpenAI Marketplace | Spend your OpenAI commitment on partner software | Eligible enterprises |
| Ultrafast | Up to 8x faster in Codex, 6x in the API, premium price | API; Pro 500 and Enterprise in ChatGPT |
Also shipped: Codex in the cloud, a refreshed Codex CLI, Codex Security Cloud, a Decisions API for fast classification, Sign in with ChatGPT across 16 partner tools, a Pro 500 plan, and a Private Intelligence program with zero data retention. The full list is in OpenAI's recap.
(As a side note, getting ready for this DevDay is the first time I remember ever, in OpenAI history, saying "this is too much stuff to launch".)
The Cheap Model Is the Real News
Most agent work does not need the frontier model. It needs a model that is good enough at tool use and cheap enough to run thousands of times a day. OpenAI says GPT-6.1 Sol matches Astra on its DeepSWE coding benchmark at roughly one-fifth the cost, per the Sol launch post. Those are vendor numbers, so the first independent read matters more, and it landed within three hours.
GPT-6.1 Sol replaces GPT-6 Sol after just 7 days. It scores 1 point below GPT-6 Astra in the Intelligence Index at less than one quarter of the Cost per Task.
The line that matters for agents is cached input at $0.10 per million tokens. An agent resends the same long context on every step: the system prompt, the tool definitions, the scratchpad. When most of your input is cache, the cache price is your bill. Here is what that does to a realistic workload.
Sol saves $7,776 a month on this workload, or $93,312 a year, if it finishes the job in the same number of steps.
If you priced an agent project on Astra this month and shelved it, reprice it on Sol. The caveat is in the fine print of the calculator: a cheaper model that takes more steps or more retries on a hard task gives some of the savings back. We covered where Astra still earns its price in GPT-6 Astra vs Claude Fable 5.1.
What to Hand Your First Dot
We have been running Grok Bot across our own sales, CRM, reporting and ops work, on top of scheduled agents that already reconcile our lead data every morning and post a standup briefing to Slack every afternoon. Dots are a different product, but the lesson carries straight over: the unit of work is a routine, not a prompt.
Most people will open their dot and type a request, the way they would into a chatbot. The better question is "what do I do every week, the same way each time, across three or four tools?" The valuable routines are the boring ones nobody does consistently by hand. And they should be handed over in a specific order.
| Stage | Hand it | Dot setting |
|---|---|---|
| 1. Read-only | Monday pipeline brief, overnight account research, numbers pulled from portals | Proactive research, which OpenAI restricts to read-only tools |
| 2. Drafts behind approval | Follow-up emails, weekly status updates, invoice reminders | Custom Rule set to Ask |
| 3. Writes to systems of record | CRM cleanup, stage updates, invoice entry | Allow, only after weeks of clean runs on a narrowly scoped connection |
| Never | Payments, password changes, anything sent in a leader's name without review | Block (OpenAI already keeps password changes with you) |
The dots launch post describes the controls that make that ladder possible. Custom Rules let you allow, require approval for, or block specific actions. Some actions, like changing a password, always stay with you. An Activity View shows what the dot did in the background. Flip the rules below and watch the same day play out differently.
Background research: read-only tools only.
- 06:10Read overnight support tickets and flagged two repeatsRead-only, done
- 07:30Drafted a reminder for an invoice 14 days overdueNeeds approval
- 09:05Moved a deal to Proposal after the call notes landedCRM updated
- 11:40Wrote the weekly pipeline update for #salesNeeds approval
- 14:15Vendor portal asked for a password resetHanded to you
Start every routine on Ask. Move it to Allow when you are bored of approving it, not before.
Dots vs Grok Bot vs Muse
Dots are the third always-on agent in seven weeks, after SpaceXAI's Grok Bot and Meta's Muse. All three give the agent its own computer and your logins.
grok bot, muse, and now dots - they may look similar, but they are not. i think openai's dots is a significantly more cohesive implementation that spans across personal and business use.
We agree with the direction of that take. Last week we argued that the personal agent and the work agent are not the same product: Muse has no answer for what happens when an employee leaves, and Grok Bot does. Dots try to be both. A personal dot on Pro and Business Premium, plus a preview of "specialist dots" with their own company identity and credentials, managed through Microsoft Agent 365. That second piece is the one to watch. Until it ships beyond pilots, treat a dot as a personal agent that happens to live in a work account.
OpenAI Is Coming for the Workspace
Put Space, Pages, collaborative slides, team tasks, @ChatGPT in Slack and plugin extensions side by side and the strategy is plain. OpenAI wants ChatGPT to be where the document lives, where meeting notes land, and where third-party software shows up, with its own agent in the middle of it.
The Marketplace is the tell. Letting an enterprise spend its OpenAI commitment on HubSpot, Salesforce or ServiceNow turns OpenAI from a vendor into a procurement channel. For a CFO, that is a real consolidation argument. For everyone else, it is a lock-in question to ask before the renewal, not after.
How DevDay 2026 Affects Your Business
- On ChatGPT Business or Enterprise: dots are off by default on Enterprise, which is correct. Turn on the beta for a small pilot group with narrow app connections. Decide who owns an agent's permissions when an employee leaves before the first dot is created.
- On Claude: nothing here is a reason to switch. Anthropic already has Managed Agents and its own workspace tools. What DevDay changes is price pressure on the mid tier, which helps everyone buying.
- Building AI into your own product: the Agents API and Bedrock Managed Agents remove a lot of plumbing. The Decisions API confirms what we found when we benchmarked 150 business decisions: most production calls are a label, not an essay, and should not cost like one.
- In the EU or UK: check availability first. Dots are limited to eligible markets, and Ultrafast runs only on US data residency and global processing, per OpenAI's Ultrafast documentation.
What We Are Telling Clients This Week
1. Reprice one shelved agent project on Sol
If cost was the blocker, run it through the calculator above. It may not be anymore.
2. Pilot one dot on one recurring job, with a named owner
Weekly status updates, invoice follow-ups, inbound triage. Something with a clear definition of done that a person checks every week.
3. Write the agent policy before you need it
Which apps an agent may connect to, what always needs approval, who reviews the activity log, and who revokes access when someone leaves. That document takes an afternoon and saves a very bad week.
Do You Need a Partner?
Not for a single dot on a single job. A motivated owner and the controls above are enough. A partner earns the fee when agents touch shared systems, client data or more than one team, or when you are deciding between OpenAI and Claude for a company-wide rollout. We run both, and that decision is most of what our OpenAI consulting work covers.