OpenAI held DevDay 2026 at Fort Mason in San Francisco on September 29, with Sam Altman’s keynote livestreamed at 10:00 AM Pacific. DevDay event page The company’s recap counts more than 20 announcements and pitches ChatGPT to developers as a way to reach “our collective 1.2B weekly users”. Two products lead the recap. The first is dots, an always-on agent with its own cloud computer. The second is GPT-6.1 Sol, which OpenAI says nearly matches GPT-6 Astra on agentic coding, computer use and professional work at a fifth of Astra’s standard input and output prices. OpenAI’s DevDay recap
Many of the agent launches do work while the user is away: a dot researching in the background, a team task fired by a Slack message, an MCP event starting an automation. Codex offers cloud tasks that run with the laptop closed, automatic code reviews and repository security scans. Our read is that what those share is a need for permissions, review steps and an audit trail, more than a need for better chat. Availability is also fragmented. The recap’s availability lines are often broader than what the help center and developer docs describe, as with dots, Ultrafast, Sites and Codex Security Cloud.
S5 Labs has not tested any of these products. Most of what follows comes from OpenAI’s launch pages, documentation and system cards, plus AWS, the MCP project and news coverage where linked; vendor-reported results are marked as such.
The recap’s announcements at a glance
| Announcement | Availability on Sept 29 |
|---|---|
| Dots: always-on agent on GPT-6 Astra with its own cloud computer | Rolling out to Pro and Business Premium in eligible markets; Enterprise beta off by default |
| Specialist dots: dots with their own identity for defined roles in a company | Preview; enterprise pilots only |
| GPT-6.1 Sol: upgrade to GPT-6 Sol at the same input and output price | API, ChatGPT Work and Codex; not yet in Chat |
| Ultrafast: premium speed tier, broadly available for GPT-6 Astra | API at low rate limits; Pro 500 and eligible Enterprise and Edu; Sol version coming |
| Private Intelligence: ZDR with Private Safety Processing, plus Private Inference | ZDR with PSP for organizations approved for ZDR; Private Inference preview “coming this fall” |
| Codex in the cloud: publishable, shareable cloud environments | Plus and up; Enterprise off by default |
| Codex CLI refresh: voice, /agents view, worktrees, cleaner UI | All plans; most pieces shipped before DevDay |
| Code Review: desktop review client plus automatic cloud reviews | GitHub generally available; GitLab in preview |
| Codex Security Cloud: repository scans and commit monitoring; recap says Daybreak Blue models are included | Research preview for workspaces with access |
| Decisions API: Luna-based classification and routing with fixed answers | Limited preview; no public docs |
| Agents API computer use: hosted browser inside the Agents API | Added to the Agents API public beta |
| Bedrock Managed Agents: Agents API harness and inference running in AWS | Preview, first announced April 28 |
| Plugin extensions: sidebar apps, panels and file viewers | Live; Free and Go on web “coming soon”; some parts desktop-only |
| Plugin Creator and submission: build help, redesigned review flow, ranking | Live |
| Sites can host plugins: workspace-private Sites with plugin access | Public beta for Business, Enterprise and Edu (Healthcare per recap); Enterprise plugins off by default |
| MCP events: plugins start automations from app events | Live, against a draft specification |
| ChatGPT Space: shared workspace replacing Library | Pro, Business, Enterprise; editing on desktop and web |
| Pages: documents shared by people and agents | Pro, Business, Enterprise; “Keep Updated” not at launch |
| Collaborative slides: multi-editor decks | Not yet; “in the coming weeks” |
| Teams and team tasks: shared tasks on a schedule or event trigger | Business and Enterprise |
| @ChatGPT in Slack and Teams: mention ChatGPT in channels and DMs | Business and Enterprise |
| Meetings plugin: meeting notes saved to Space | macOS desktop beta for Pro and Business; Enterprise coming soon |
| Shareable profiles: one page collecting a person’s Sites and plugins | Free through Enterprise, per the recap |
| Sign in with ChatGPT: identity plus plan usage in partner apps | Identity global per the recap, commercial partners in a limited trial; plan usage for Plus and Pro |
| Pro 500: $500/month plan with Ultrafast | Available now; Pro 200 reopened with a lower allowance |
| OpenAI Marketplace: apply part of an enterprise commitment to eligible products from 32 partners | Enterprise customers contact OpenAI to confirm eligibility |
Sources for the table are the recap and the product pages cited below. OpenAI’s DevDay page in the ChatGPT documentation adds four items the recap leaves out: ChatGPT Work tasks that sync across devices, a Browser Annotation API for websites, company-managed workspace connections, and Agent Security policies for tool, file and network access. DevDay 2026 in the ChatGPT docs
Dots: action controls outpace memory controls
A dot is a persistent agent running on GPT-6 Astra in the cloud, with its own computer and browser, that keeps working when the user’s machine is off. A comparable product is SpaceXAI’s Grok Bot. Introducing dots · Meet dots
Who can get one and what it costs for now
Dots began rolling out on September 29 to Pro and Business Premium subscribers, and rollout can take days. The help center excludes Pro users in the European Economic Area, Switzerland and the UK, and dots are not available to anyone under 18. Texting is a limited beta for some US Pro users, though the launch post still calls it coming soon. Microsoft Teams access is an invite-only alpha per the Enterprise admin guide, while the launch post lists Teams without qualification. A dot cannot be created on mobile. Enterprise, Edu and Healthcare get a beta that an administrator must switch on. For Enterprise, that beta has no data or inference residency, and it excludes FedRAMP workspaces, workspaces using EKM, and workspaces with UAE inference residency. Getting started with your dot · Dots privacy FAQ · Enterprise admin guide · Cloud and local access
The first dot is included in Pro and Business Premium at no extra cost. The launch post says conversations with it do not count toward usage limits, and that tasks it starts in Codex or ChatGPT Work count as usual. It also says plans include extended limits for deeper work in the first month. The help center says dots usage won’t count toward plan allowances “for the next month”, does not mention Codex or Work tasks, and says usage terms will follow after that period. The cost of a dot doing real delegated work is therefore unknown until at least late October. Dots launch post · Getting started with your dot
The safety architecture is strongest on the action side
Proactive research, the background work a dot does unprompted, runs on read-only tools. OpenAI says the limits are enforced in code: those tasks cannot directly send messages to other people, change content in connected apps or control a browser or desktop. Any follow-up action goes through the normal rules. Before a dot sends an email or edits a file, a separate reviewer called Auto-review checks the planned steps against the user’s instructions, Custom Rules and safety requirements. For supported sign-ins, secure sign-in pauses the model while the person fills a login form, so the password never enters model context. Changing a password or moving money between accounts is handed back to the person. How OpenAI builds safety into dots
Two caveats come from OpenAI’s own pages. The Auto-review documentation the launch post links to is Codex documentation, and it says Auto-review “is not a deterministic security guarantee”. The GPT-6 Astra system card’s dots appendix reports no scored prompt-injection successes across 100 bulk rollouts of 500 simulated emails each. It also reports unwanted persistence after a warning in 17.4% of rollouts at the maximum reasoning budget, in a low-stakes test run without system-level controls. Moderate scope violations rose from 8.6% to 19.7% of samples as intervening tasks went from five to ten. OpenAI describes “known vulnerabilities” that it judges hard enough to exploit to justify launch. Auto-review docs · GPT-6 Astra system card, dots appendix
Memory controls are all-or-nothing
A dot forms memories from connected apps proactively. Users currently cannot view, delete or directly modify individual memories, including details that arrived through plugins. Disconnecting an app stops new access but does not remove what the dot already learned. Turning off ChatGPT Memory stops the sharing but does not delete what the dot has already received. The only deletion path is deleting the dot itself, which the help center calls a reset. That removes its conversations, memories and scheduled tasks and cannot be undone. Dots privacy FAQ · Control your dot
Settle that asymmetry before connecting a work mailbox or CRM: the action side has confirmations, handoffs and a reviewer, while the memory side offers only deleting the dot. Anthropic’s Dreaming feature has a review mode where memory changes wait for approval, covered in our Code with Claude 2026 article; dots have no visible equivalent at launch.
GPT-6.1 Sol keeps GPT-6 Sol’s input and output price
GPT-6.1 Sol is in the API as gpt-6.1-sol. Standard pricing per million tokens is $2 input, $0.10 cached input, $2.50 cache writes and $10 output for prompts up to 272K input tokens. Above that, input and cache rates double and output goes to 1.5x for the whole request. Model page · API changelog
Against GPT-6 Sol, only cached input got cheaper
OpenAI’s framing is near-Astra intelligence at “a fifth of its standard input and output token prices”. That is exact on input and output against Astra’s $10 and $50, and cached input is one-tenth of Astra’s $1. But GPT-6 Sol, released a week earlier, already cost $2 in and $10 out. Against the model it upgrades, the only list-price change is cached input, halved from $0.20 to $0.10; cache writes stay at 1.25x input. DevDay recap · Pricing page · Our GPT-6 Sol and Luna article
The saving lands on agents that re-read a large prefix. Take a 200,000-token prefix written once and read 49 more times, priced from list rates, excluding output and assuming every read hits the cache. It costs $1.48 on GPT-6.1 Sol, $2.46 on GPT-6 Sol, $12.30 on Astra and $0.123 on Luna. Uncached, the same 50 requests cost $20 on either Sol. Prompt caching guide
Migrating from GPT-6 Sol: endpoints and reasoning effort
GPT-6.1 Sol does not support the none or minimal reasoning efforts, where GPT-6 Sol offered none. Tool calling requires the Responses API, whereas GPT-6 Sol allowed function calling on Chat Completions with effort set to none. A routing setup that calls functions on Chat Completions with reasoning off would therefore have to move to the Responses API and use at least low effort to adopt GPT-6.1 Sol. GPT-6 Sol and Luna still accept none. EU data residency is supported, as on GPT-6 Sol, but not with Fast mode. Model page
Vendor benchmarks and their caveats
OpenAI reports that GPT-6.1 Sol matches Astra on DeepSWE v1.1 at about a fifth of the cost and beats GPT-6 Sol’s best score by 6.4 points. It also reports $5.47 per task on Terminal-Bench Science at max effort, against $23.80 for Astra. These are OpenAI’s numbers, and the post’s text gives few absolute scores; OpenAI itself says Astra still leads the hardest science work at 68.1% on Terminal-Bench Science. GPT-6.1 Sol launch post
The system card addendum adds context the launch post leaves out. Under its Preparedness Framework, OpenAI treats GPT-6.1 Sol as Critical in cybersecurity and High for biological and chemical capability, the same treatment as Astra. These are capability thresholds, and the model ships with Astra’s safeguards stack. On a coding-deception test built to elicit dishonest behavior, its misrepresentation rate is 1.50%, against 0.51% for Astra and 1.30% for GPT-6 Sol. Unwanted persistence after warnings appeared in 23.5% of rollouts against 17.4% for Astra, in tests run without system-level anti-circumvention controls. System card addendum
In ChatGPT, GPT-6.1 Sol is available to Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex, not yet in Chat. Enterprise and Edu keep it off until an administrator enables it, and Free and Go are not included at launch. Models
Ultrafast: 6x Standard, broadly available only for Astra
Ultrafast is a service tier, not a model. In the API it is broadly available for gpt-6-astra, with preview access for GPT-5.6 Sol through OpenAI account teams. Astra Ultrafast costs $60 input, $6 cached, $75 cache writes and $300 output per million tokens for prompts up to 272K tokens. That is exactly 6x Astra Standard on every column, and 30x GPT-6.1 Sol’s standard output price. It is open to all API users at low default limits (500,000 tokens per minute on tiers 1 to 3). It supports US data residency and global processing only, with no EU residency. Ultrafast guide · Pricing page
The speed claims disagree with each other. The recap says up to 8x faster token generation (300 tokens per second) in Codex and up to 6x in the API, while the API guide’s subtitle says up to 8x. In ChatGPT, Ultrafast draws included usage at 8x the Standard rate and credits at 6x, and is limited to Pro 500 and eligible Enterprise and Edu plans. At launch, buying credits on Pro 100 or Pro 200 does not unlock it. GPT-6.1 Sol Ultrafast has no price, rate limit or date beyond “in the coming days”. Speed · Pro tiers help article · GPT-6.1 Sol launch post
Developer platform: computer use, Decisions API, Bedrock and privacy
Agents API computer use is a hosted browser
The September 29 changelog records one Agents API change: computer use, which ships inside the Agents API public beta. Multi-agent, tool search and compaction were already in the September 10 public beta covered in our Agents API article. Computer use means browser use in an OpenAI-hosted session. Every documented example uses gpt-6-astra, and no Agents API supported-model list is published (model pages list computer use only for the Responses API). Computer use guide · API changelog
The browser needs the user’s approval before reaching each new website origin. Sign-in is handled by the developer’s application through a dedicated event that keeps credentials out of model input; passkey and QR-code sign-ins are not supported. The design does not enforce per-action confirmation. OpenAI’s doc says an application that must guarantee confirmation before purchases or destructive changes should either restrict the hosted browser to resources that cannot perform them or use a browser runtime the developer controls. Computer use guide
The recap also lists it “in Codex and ChatGPT Work on Pro 500 and Enterprise”. No Agents API documentation we could find backs that; the wording matches the Ultrafast entry, and the Agents API still needs an API key. Codex and ChatGPT Work separately document a desktop Computer Use plugin, which is a different feature. Agents API overview
Decisions API has a paragraph and a preview
The Decisions API focuses “Luna’s intelligence” on a fixed set of user-defined questions with finite answers, taking text or images and returning answers for classification, routing or choosing an agent’s next action. It is in limited preview with broad release “in the coming days”. The Decoder reports OpenAI’s claim that it makes decisions ten times faster than GPT-6 Luna through the regular API. No OpenAI page we could open gives a model ID, endpoint, schema, price or rate limit. DevDay recap · The Decoder
Bedrock Managed Agents and Private Intelligence
Bedrock Managed Agents is not a DevDay launch: AWS announced it on April 28 in limited preview, and it is still labeled Preview. OpenAI’s comparison page draws the line. A self-hosted sandbox on OpenAI’s Agents API changes only where commands run, while harness and inference stay with OpenAI. Managed Agents moves both into Bedrock, authenticated with AWS IAM credentials, and AWS says each agent gets its own AWS identity. We found no AWS pricing or supported-region list. OpenAI’s Bedrock comparison · AWS product page · AWS April 28 announcement
Private Intelligence has two halves. Zero Data Retention with Private Safety Processing is documented for organizations already approved for Zero Data Retention. Sampled or classifier-referred prompts and responses are encrypted into the customer’s own cloud bucket and kept at least 30 days. By OpenAI’s design, they are decrypted only inside a hardware-attested runtime that disables human access. “Zero retention” here means OpenAI does not keep the content; the customer’s bucket does. Private Inference is a preview “coming this fall” with one sentence of description. ZDR with Private Safety Processing
Our September 11 article noted that the Agents API supported US-only residency and no Zero Data Retention. That is still what the overview says. The data-controls page marks /v1/agents as ineligible for Zero Data Retention and says its endpoint limitations still apply to ZDR with Private Safety Processing. Agents API overview · Data controls
Codex: cloud environments are new, much else shipped earlier
Cloud environments
Codex now prepares a cloud environment conversationally: it inspects the repositories, installs dependencies, tests the setup, and can record an install script and a start skill. Publishing captures the prepared filesystem. Each new task starts from it in an isolated workspace with its own working state, and the published environment changes only when someone edits and republishes it. Enterprise workspaces default to off. Computer and browser use, GitLab and self-hosted GitHub Enterprise Server are not yet supported in the new environments (OpenAI lists them as roadmap items), though GitLab reviews and tasks run in beta through the older flow. Code Review and the GitHub and Linear integrations still run on “Codex Cloud (Legacy)”, which OpenAI plans to deprecate. Cloud environments · Admin rollout guide · GitLab guide
CLI, Code Review and Security Cloud
The recap’s CLI features mostly predate the event: the agents overview arrived in early September, experimental worktrees on September 9, voice and worktrees became defaults on September 22, and fullscreen transcripts became the default on September 25. Version 0.159.1, published on GitHub the afternoon of the keynote, makes GPT-6.1 Sol the default model in the bundled catalog. Codex releases · ChatGPT and Codex changelog
Code Review on GitHub is generally available; GitLab merge-request support is in preview (the GitLab guide calls it beta). Automatic GitHub reviews flag only P0 and P1 issues, and a chat review never posts, approves or merges on its own. Code Review docs · GitHub integration
For Security Cloud, the recap promises scheduled scans, duplicate removal, prepared fixes and Daybreak Blue models “without a separate Daybreak application” for all Pro, Business, Enterprise and Edu users. The docs describe a one-off Repository scan or Commit changes monitoring, validation by reproducing findings in a container, and a proposed patch a human turns into a draft pull request. The Security Cloud pages document no recurring schedule, deduplication or Security Cloud pricing, and do not explain how Daybreak Blue is applied. The Codex pricing page separately says Daybreak access requires Trusted Access for Cyber approval. The product is a research preview for workspaces with access, and the pricing matrix lists Codex Security under a single Enterprise and Education column. For what Daybreak Blue is, see our Daybreak coverage. Codex Security · Security Cloud setup · Security Cloud FAQ · Codex pricing
ChatGPT as a distribution platform
Plugin extensions, submission and Sites
Extensions give a plugin a sidebar home, a panel beside the conversation, file viewers, composer mentions and rich forms. They are built by adding metadata to an MCP server’s tools with OpenAI’s new MCP Extensions SDKs (version 0.1.0, created on launch day), on top of the open MCP Apps standard. The spec’s launch-day table (“expected support”) makes file viewers and composer mentions desktop-only, and the extensions docs say web extensions for Free and Go are coming soon. Plugin Extensions docs · Extensions spec
Submission is a ZIP upload, automated checks, then review. A plugin with an MCP server needs exactly five positive and three negative test cases and a walkthrough video for its initial review. Feedback arrives by email with no published turnaround. Plugins appear on the directory’s main pages only if OpenAI selects them, and developers cannot request it. Commerce is currently physical goods only, digital subscriptions cannot be sold, and no revenue share is published. Submission docs · Review requirements · Plugin guidelines
Sites is in public beta. It can host plugins only in workspace-private Sites in workspaces where the feature is enabled, with each visitor connecting their own accounts, and Enterprise starts with plugins off. Sites has no data or inference residency at launch, and OpenAI says not to use it for Protected Health Information or payment-card data, which sits oddly beside the recap’s Healthcare listing. Sites docs · Sites administration
MCP events build on a draft
ChatGPT now subscribes to events from a plugin’s MCP server and runs a task when one arrives. OpenAI’s recap calls the spec “proposed” and its docs call it a draft. The working group’s incubation repository is labeled experimental and says its contents are not official MCP specifications. The group’s charter lists the events SEP as “Ideating” with an end-of-April target. ChatGPT supports only webhook delivery, signed with Standard Webhooks, plus callback verification. Polling, streaming and the draft’s gap and terminated notifications are not supported. It also requires MCP protocol version 2026-07-28, the revision that removed protocol-level sessions and the initialize handshake. Building on this today means building against a wire format that can change. MCP events docs · Triggers and Events Charter · Incubation repository · MCP 2026-07-28 changes
Sign in with ChatGPT: whose allowance pays
Sign in with ChatGPT is two things. Identity is standard OAuth and OpenID Connect; the recap says it is available globally, though commercial partners are still in a limited trial with a waitlist. Plan usage lets a Plus or Pro subscriber’s plan pay for a partner app’s AI requests without an API key. The recap’s 16 partners appear to be the 12 plan-usage apps in OpenAI’s directory (Devin, Notion, Vercel, Warp and others; one, Lovable, is marked coming soon) plus four open-source integrations. That sum is our inference. SIWC quickstart · Partner directory
The help center is precise about the user side. Eligible requests count toward the plan’s ChatGPT Work and Codex usage. A per-app weekly cap is a percentage of overall usage, not a separate pool. Usage rates may differ between an app and ChatGPT, and usage settings do not show details of what you did in the app. Credit use after the plan limit is off by default. If automatic credit purchases are on, continued usage can trigger automatic charges, and OpenAI says you may not get a separate notice when the plan limit is reached. For developers, plan usage is open to open-source partners and selected private clients, with a published list of preview limitations. OpenAI does not notify an app when a user disconnects, and no fee terms are published. Using your ChatGPT plan in other apps · Preview limitations · Errors and recovery
Collaboration: announcements run ahead of what ships
Space replaces Library for Pro, Business and Enterprise on desktop and web, with finding, reading and sharing on mobile. Mobile creation and editing are coming. Pages are documents where people and agents edit together. The marketing page says they can stay updated from connected tools, while the docs say “Keep Updated isn’t available at launch” and point to scheduled tasks set up in chat for recurring updates. Collaborative slides are “coming soon” on the Space page and “in the coming weeks” in the recap. Space · Work with agents in Space · DevDay recap
The recap lists teams and team tasks for Business and Enterprise. Team tasks run in the cloud with a team service account and its connected apps, on a schedule or on an event such as a Slack message, and bill to workspace credits. Three governance points stand out. Anyone who joins a team sees all earlier runs and generated files. A team connection may reach information a member’s own account cannot. And neither pausing nor deleting a task should be relied on to interrupt a run in progress. Teams and Team Tasks
@ChatGPT in Slack and Teams answers mentions using tools an admin connected and has no memory at launch; OpenAI says teammates can join the conversation without an individual ChatGPT license. The Meetings plugin is a macOS desktop beta for Pro and Business with Enterprise “coming soon”, and OpenAI says audio is deleted once notes are ready. Slack and Teams docs · DevDay recap
Plans and procurement: Pro 200 cut matters more than Pro 500
Pro now has three monthly tiers: Pro 100 at $100, Pro 200 at $200 and Pro 500 at $500. Pro 500 includes Ultrafast and, per the recap, 25 times the Plus allowance; the help article states no multiplier. Pro 200, which TechCrunch reported OpenAI paused for new sign-ups on September 10, is open again at the same price with a lower allowance. Pro 200 subscribers who were active at the eligibility cutoff, or in the seven days before it, keep the previous allowance through October 29, 2026, as long as they stay subscribed. Pro tiers help article · TechCrunch on the pause
Thibault Sottiaux of OpenAI framed the change as raising the floor across Plus and Pro. In a post on X, he put the new relative sizes at Plus 1x, Pro 100 5x and Pro 200 10x. Existing plans, he said, keep the 20x multiplier “for a bit” and “also receive a lot of additional credits”. The Next Web reports a one-time grant worth $2,500 that expires at the end of the year; we did not find that figure on an OpenAI page. Sottiaux on X · The Next Web
Arithmetic from those ratios and list prices: measured against Plus, every Plus and Pro tier now works out to $20 per Plus-sized unit, where the old Pro 200 was $10. Because the Plus floor itself is rising, this compares multipliers, not absolute usage. Pro 500 carries no volume discount; what it buys is headroom and Ultrafast. And because Ultrafast draws included usage at 8x, a Pro 500 subscriber running everything in Ultrafast would get roughly 3x a Plus allowance of work (25 divided by 8). That assumes both multipliers apply to the same unit, which OpenAI does not state. Codex pricing
The OpenAI Marketplace lets eligible enterprise customers apply part of an existing OpenAI commitment to eligible products from 32 partners, including Figma, Adobe, Salesforce, ServiceNow, Harvey, Palo Alto Networks, CrowdStrike and Baseten. The customer contracts with and pays the partner directly, and OpenAI reconciles the commitment. Eligibility is per product rather than per partner, and how much of a commitment can be applied is not published. Marketplace
What to test first and what to watch
GPT-6.1 Sol on your own tasks. Its list price matches GPT-6 Sol except cached input, which is half. For teams on Astra, the question is whether quality on your workload is close enough to move down; for teams on GPT-6 Sol, whether the effort and endpoint changes fit. Run a representative set at a fixed effort and log cost per completed and checked task. If prompts share a large prefix, measure the cache hit rate over a real session; against GPT-6 Sol, the halved cached-read price is the only list-price saving.
Ultrafast with a stopwatch. On Astra, run the same Codex task in Standard and Ultrafast (on Pro 500 or an eligible Enterprise or Edu workspace; the API is the alternative) and compare wall-clock time against the usage dashboard. OpenAI’s speed figure measures token generation, not task completion.
Agents API computer use against the hard cases. Multi-origin redirects, a site that needs a passkey, and a task that reaches a purchase or delete button. Origin approval does not stop that click; confirm the browser restriction or run your own runtime.
Dots with a sacrificial account. It still has to be a paid Pro or Business Premium account in an eligible market. Connect a read-mostly source, keep any “take action without asking” rules narrow, and note what the dot remembers, because you cannot edit it afterwards. Decide whether you would delete the dot before a work mailbox goes in. Note when the one-month terms end: about October 29 if counted from launch, though OpenAI gives no date and its pages differ on what ends.
Team tasks after a membership review. Before the first event-triggered task, review who is on the team and what its connected accounts can reach. Anyone who joins sees all earlier runs and generated files, and a connection may reach information a member’s own account cannot.
Still to watch: Decisions API documentation and pricing, GPT-6.1 Sol Ultrafast pricing, and the post-October dots usage terms. Also open: whether the Agents API gains ZDR or EU residency, whether the MCP events SEP is accepted or changes shape, and whether Security Cloud availability settles between the recap’s four plans and the pricing matrix’s single Enterprise and Education column.
Separately, CNBC reported on September 28 that OpenAI would not release GPT-6.1 Astra because it did not adequately meet the company’s safety standards. Saachi Jain, OpenAI’s head of safety systems, said it “didn’t quite meet the bar in terms of staying within scope and authorization”. The DevDay pages we reviewed do not mention it, so the top of OpenAI’s lineup remains the Astra that shipped on September 3. CNBC
Sources
OpenAI announcements
- DevDay 2026 Recap and OpenAI DevDay 2026 event page — OpenAI
- Introducing dots — OpenAI
- How we build safety, security, and privacy into dots — OpenAI
- Introducing GPT-6.1 Sol — OpenAI
- OpenAI Marketplace — OpenAI
- ChatGPT Space — OpenAI
System cards
- Addendum to GPT-6 Astra System Card: GPT-6.1 Sol — OpenAI
- GPT-6 Astra System Card (change log and dots appendix) — OpenAI
Developer documentation
- API changelog, GPT-6.1 Sol model page, Pricing, Prompt caching, Ultrafast mode — OpenAI Developers
- Agents API overview, Agents API computer use, Bedrock Managed Agents, ZDR with Private Safety Processing, Data controls in the OpenAI platform — OpenAI Developers
- Plugin Extensions, MCP Events, Upload and submit your plugin, Remote MCP server review requirements, Plugin guidelines — OpenAI Developers
- Sign in with ChatGPT quickstart, Preview limitations, Errors and recovery — OpenAI Developers
- MCP Extensions specification and Codex CLI releases — OpenAI on GitHub
ChatGPT and Codex documentation
- DevDay 2026, Models, Speed, Pricing, ChatGPT & Codex changelog — ChatGPT Learn
- Meet dots, Control your dot, Manage dots permissions and capabilities, Local computer access for Work Cloud and dots, Auto-review — ChatGPT Learn
- Cloud environments, Admin rollout guide, Code Review, Review GitHub pull requests with Codex, Review GitLab merge requests with Codex, Codex Security, Codex Security Cloud setup, Codex Security Cloud FAQ, Computer Use — ChatGPT Learn
- Sign in with ChatGPT, Sites, Sites administration, Work with agents in Space, Set up and manage teams and Team Tasks, Set up and manage @ChatGPT in Slack and Microsoft Teams — ChatGPT Learn
Help Center
- Getting started with your dot, Dots privacy, security, and safety FAQs, About ChatGPT Pro tiers, Using your ChatGPT plan in other apps and sites — OpenAI Help Center
Other
- Amazon Bedrock Managed Agents, powered by OpenAI (Preview) and Amazon Bedrock now offers OpenAI models, Codex, and Managed Agents (Limited Preview) — AWS
- Triggers and Events Charter, experimental-ext-triggers-events and Key Changes in 2026-07-28 — Model Context Protocol
- Tibo (Thibault Sottiaux) on the new Pro 200 plan — X
- OpenAI puts Pro subscriptions on hold due to Astra demand — TechCrunch
- OpenAI halves Pro 200 usage and launches a $500 ChatGPT plan at DevDay — The Next Web
- OpenAI expands Codex and its API at DevDay with security scans, a Decisions API, and Ultrafast — The Decoder
- OpenAI abandons plan to release upcoming model as safety concerns escalate — CNBC