OpenAI DevDay 2026: Dots, Ultrafast, GPT-6.1 Sol, and Space

Matthew Berman attended OpenAI DevDay 2026 and came back with a rapid tour of the announcements: an always-on agent, much faster Astra responses, a new Pro tier, a cheaper near-frontier model, more Codex surfaces, and a shared place to work in ChatGPT. The demo energy is real. The useful question is which parts you can actually use, in which region, and at what cost.

Watch Matthew Berman's Recap

Credit: Matthew Berman's video, published 29 September 2026. He attended DevDay and tested Astra Ultrafast in early access. Product details below are checked against OpenAI's public documentation as of 30 September 2026.

What Is Available, and What Needs Checking?

Dots: A Persistent Agent With Real Access Boundaries

At 00:17, Matthew describes Dots as OpenAI's answer to personal agents that can keep working between conversations. OpenAI's Dots guide confirms an always-on agent in ChatGPT, powered by Astra, with its own cloud computer, connected apps, memory, scheduled work, and progress for the user to review. Local computer access is optional and off by default; if enabled, a dot can create Work or Codex tasks on that computer.

There is an important regional detail for readers in Portugal: the initial personal Pro rollout excludes the European Economic Area, Switzerland, and the UK. Business Premium is available in supported ChatGPT regions; Enterprise, Edu, and Healthcare beta access depends on an administrator enabling it. Rollout is gradual, and the first dot is included for eligible plans, with an allowance for deeper work. That is not unlimited autonomous labor.

Connected apps can supply context proactively, so start with the least access that supports one useful task. Check the dot's rules for approvals and handoffs, and review anything involving purchases, messages, client data, or file changes. OpenAI explicitly says a dot can make mistakes even when following custom rules.

Ultrafast and Pro 500: Speed Has Two Price Tags

Matthew's favorite launch is Astra Ultrafast. In his side-by-side rocket-building demo, the Ultrafast run reaches an output while Standard is still working. OpenAI's API guide says the service tier can deliver up to eight times faster output than Standard and recommends persistent WebSockets for tool-heavy agents, where network and tool overhead can otherwise erase some of the gain. API access to Astra Ultrafast is broadly available at low rate limits; higher limits require a request.

The price is not a footnote. In OpenAI's short-context API rate card, Astra Standard is $10 per million input tokens and $50 per million output tokens; Ultrafast is $60 and $300 respectively. Matthew says he spent about $1,000 in 90 minutes of early-access use. That is his workload, not the expected cost for everyone. For a production decision, compare elapsed time and dollars per successful job, including tool calls, retries, and human review.

Pro 500 is a separate $500-per-month ChatGPT subscription that includes Ultrafast among personal Pro plans. Pro 100 and Pro 200 do not include it, and buying credits on those tiers does not unlock it at launch. Ultrafast consumes included usage first, then available credits. OpenAI also says new Pro 200 subscriptions have a lower allowance than previously offered; eligible existing subscribers keep their earlier allowance until 29 October 2026. Check your account's actual limits before changing plans. Matthew's 25x and 10x comparisons are his reading of the launch, not a substitute for the current plan screen.

GPT-6.1 Sol: Test the Cost per Good Result

The GPT-6.1 Sol model page positions it near Astra for complex coding, computer use, and professional work at a lower cost. For standard, short-context API requests, the published rate is $2 per million input tokens, $0.10 per million cached input tokens, and $10 per million output tokens. Astra Standard is $10 input and $50 output. That is a fivefold difference in the base input and output rates, not a promise that every entire task will be five times cheaper.

Matthew points to benchmark charts where Sol approaches Astra. He also notes Astra can still lead at the highest computer-use settings. His theory that post-training has been "solved" is an inference, not a disclosed explanation of how OpenAI built the model. The practical move is to run the same tasks through both at matching reasoning effort, record success and correction rates, then compare full cost and wall-clock time. A cheaper token is only useful if the job still gets done.

Codex: Cloud Work, Code Review, and Security

The video groups Codex Security Cloud, cloud coding, a refreshed CLI with voice control, and code review into a quick tour. The timeline matters: Codex Security entered research preview in March 2026, and cloud-based Codex work also predates this DevDay. Matthew is reacting to their expanded prominence and interface, not their first existence.

Codex Security Cloud connects to authorized GitHub repositories, builds an editable threat model, investigates candidate vulnerabilities, attempts isolated validation, and proposes patches for review. It does not automatically modify code. OpenAI says Cloud scans use token-based billing, with opt-in before paid usage for existing customers. A small team should trial it on one lower-risk repository, inspect false positives and the proposed patches, and keep its normal code-review and test process.

Matthew also shows voice control and a refreshed review flow. Treat these as creator-observed interface demonstrations: exact availability can vary by plan, app version, and workspace settings. The durable principle is to let cloud tasks run when you do not need a laptop open, then require tests and a person to review the diff before merging.

Decisions, Plugins, and Sign in With ChatGPT

At 08:19, Matthew describes the Decisions API as a preview for fast, constrained choices. He compares it with Jev and predicts Jev may remain faster; that is a prediction, not a head-to-head measurement in this video. We could not verify a public Decisions API implementation guide or price sheet on 30 September. Prototype the decision schema and evaluation set, but do not assume production access or economics yet.

OpenAI's plugin review guide confirms a route to publish tools in the shared ChatGPT and Codex directory. Approval and publication are separate steps. Publishing enables a direct listing and name search, while prominent placement or proactive suggestions are selective and not guaranteed. This matters if your business model assumes an agent will discover your tool on its own.

"Bring your tokens" needs similar precision. Sign in with ChatGPT offers identity sign-in and, separately, eligible plan-backed AI requests in participating apps. Commercial sign-in is in a limited partner trial; plan usage is available to open-source partners and selected private clients. A user logging in does not automatically transfer an unlimited subscription allowance to every third-party app. Builders should verify scopes, eligibility, limits, and fallback billing before promising a free core product.

Space Is for Pages Today, More Formats Later

Matthew ends with ChatGPT Space, imagining a shared home for teammates and agents. The official Space guide is more specific: eligible Pro, Business, and Enterprise users can create and edit pages, organize them with folders and subpages, and collaborate. Space replaces Library for accounts that have it, while Projects remain separate. Slides and Sheets are coming soon, not part of the launch editing surface.

Web and desktop support page editing; mobile supports finding, reading, and sharing, but not editing. Shared pages may contain copied or summarized information from connected sources, so a reader can see that material even if they cannot open the original source. Review the page contents and inherited sharing permissions before putting client information into a team space.

What I Would Test First

1. Model value: run ten real coding or research tasks on GPT-6.1 Sol and Astra with the same inputs and reasoning settings. Track completed jobs, edits, total spend, and time.
2. Speed value: use one latency-sensitive task to compare Astra Standard and Ultrafast. Include tool and network time, not just token output speed, and set a spend cap.
3. Agent access: if Dots is available to your plan and region, connect one low-risk app and give it one recurring job. Inspect approvals, memory, and completed results before expanding access.
4. Collaboration: pilot one Space page for a project brief or review log, verify who can read it, and wait for official Slides and Sheets availability before designing a workflow around them.

The theme running through DevDay is convergence: agents can remember, act, write code, call tools, and move work into shared documents. The best adoption test is still modest. Pick one task, one measured outcome, and a clear human review point.

Video Chapters

Sources and Link Map

• Creator and event: Matthew Berman's full recap and official DevDay event page.
• Dots and plans: Dots access, controls, and limits; Pro 100, 200, and 500.
• Models and speed: Ultrafast guide, API pricing, and GPT-6.1 Sol model page.
• Codex: March research-preview announcement and current Security Cloud guide.
• Apps and collaboration: plugin review and discovery, Sign in with ChatGPT, and Space availability and sharing.

For the founder's view of the same event, read our DevDay business-opportunities analysis. It focuses on plugins, the trigger-to-feedback workflow, and two startup ideas rather than this video tour.