See pricing
OpenAI Dev Day 2026: The releases that actually matter
How I AIAI

OpenAI Dev Day 2026: The releases that actually matter

Claire Vo sorts a keynote packed with announcements into what to try first. Spaces and Sites may matter most for teams, and one 30-minute demo of ultra fast mode cost her about $97.

September 30, 2026 · 24 min listen · 9 min read
0:00
–:––

Context

Claire Vo, founder of ChatPRD and host of How I AI, spent the day at OpenAI's DevDay in San Francisco. The keynote announced something new every two or three minutes, so she sorts the releases by what a PM, a team lead, or a developer should pay attention to. She also tested some of them early, so these are first impressions, not full reviews. It matters to PMs because several releases (a persistent agent, a shared workspace, and cheap fast model options) change what an AI product can be and how teams collaborate around it.

The Big Idea

Of the many DevDay releases, three things are worth your attention: Dots (an always-on agent), Spaces and Sites (shared work between humans and agents), and fast model options that make real-time AI products possible.

Claire's filter is practical: what fixes a real team problem today, and what hints at products you could not build before. Everything else can wait.

Key Insights

Dots: strong work, rough structure

Dots are an always-on, long-running agent. Each user on key plans (Pro, Business, Enterprise) gets a single Dot, which can work across ChatGPT, Codex, Slack, its own cloud computer, and your local computer.

  • What works: the work itself. Claire had her Dot shop for shoes, write code, and flag a conflict between her kids' swim classes and another plan, then offer to look for November options on its own.
  • What's rough: information architecture. It is confusing what belongs in a Dot thread, a chat thread, or a Codex thread.
  • Unusual interaction: you can call your Dot and it rings and picks up. Claire is unsure how that fits with live voice and transcription.
  • Her verdict: she is holding the full review until the product settles, since the OpenAI team is changing it by the hour.

Proactivity is how agents win attention

Claire expects a Dot to earn her attention by pushing useful things to her, using her chat and Codex memories. She still prefers hiring one small agent per job, the way she would hire a person, so she keeps using her Grok bots for work. She is not a "one agent to rule them all" person.

Spaces may be the underhyped release

Spaces is a place in ChatGPT to work on documents, artifacts, and slides with both humans and agents, with sharing and data controls built in.

  • Why it matters: companies have been waiting for an AI-native document and slide platform where agents can edit, not just generate. Claire says AI tools are great at making slides but weak at editing and collaborating on them.
  • Example: her Dot created a Space called "your personal scratch pad", a running list of to-dos, decisions, and notes. It matched a collaborative doc she had already built for herself so Codex, Grok, and Muse could share one to-do list.
  • How it works: you can @mention ChatGPT in a doc to get updates to it.
  • Open question: whether companies will move work into OpenAI-hosted docs, or stay with sources of truth like Slack, Google Drive, and Dropbox.

Sites plus plugins answers vibe-coded sprawl

Sites is OpenAI's built-in app builder that you can deploy and share, with sign-in through ChatGPT. The key addition is that you can now bundle connectors and plugins into a site.

  • Example: build a dashboard on data from a Snowflake plugin. When a teammate opens it, they use their own connector. If they have access to the data, they see the report. If not, they see nothing.
  • Why it matters: Claire says the top problem she hears from enterprises is what to do with vibe-coded internal tools, prototypes, and reports. Sites with plugins is OpenAI's answer.

Sol is the cheap, fast workhorse

GPT-6.1 Sol is described as nearly Astra-level intelligence at Sol speed and low cost.

  • Sol: $2 per million input tokens and $10 per million output tokens.
  • Astra: $10 per million input tokens and $50 per million output tokens.
  • Her use: Sol is her middle-of-the-line model for most engineering tasks. She prefers Astra most and will pay for it, but would pick Sol if she were cost conscious.

Decisions API adds vision

The Decisions API runs Luna-level intelligence over a predefined set of answers and returns a decision very fast. Claire sees it as a rival to Jev, a fast decisions model she has been covering on the show, but this one has vision.

  • Thumbnail demo: it went through 100 frames of a roughly 40 minute episode in under 10 seconds and found the ones where nobody looked awkward. Before, she scrubbed the video by hand.
  • Hot dog demo: it labeled a hot dog photo with 100% confidence in under a second.
  • Status: not fully released yet. She plans a head-to-head eval against Jev once it is.

Ultra fast makes real-time apps possible

Codex models run in normal mode, fast mode (2x as fast), or ultra fast (8x as fast at 6x the cost). GPT-6 Astra ultra fast is available in Codex and the API.

  • The Other Pencil: a collaborative sketch pad. Claire draws and Astra draws matching SVGs right behind her, such as turning her smiley circle into a chicken.
  • Little Starship: a 3D game with three characters (Pip, Momo, Sprout). A prompt like "build a bathroom" or "go to a tropical island" redraws the world in a couple of seconds. It can even remove gravity.
  • Caveats: latency is still a bit slow for a true real-time feel, and the Little Starship demo cost about $97 for 30 minutes.

Mental Models & Frameworks

Pick a platform on enterprise criteria

Claire frames the Claude artifacts versus OpenAI Sites contest as a fight over where the enterprise source of truth lives. She says you win on what companies care about:

  • Easy access to data
  • Governance and controls
  • Permissions
  • Ease of sharing
  • Quality of the artifacts

Use this list to judge any AI collaboration tool, not just these two.

Think in real-time intelligence

Claire's thought exercise for ultra fast: if frontier-level intelligence ran at near real-time speed, what would the user experience look like? She argues that 8x, 10x, or 20x speed opens products you could not build before, such as live drawing partners and games that rewrite themselves. Treat it as a design question first and a cost question second.

Trade-offs & Nuance

Many agents versus one agent

Dots are pitched as a replica or delegate of you, with one master Dot that is proactive, kicks off Codex threads, and does work on your behalf. Claire prefers defining micro agents for specific jobs, like hiring someone for a role. A single agent is simpler to adopt, but it concentrates confusion about which surface does what.

Speed versus cost

Ultra fast shows what near real-time frontier AI feels like, but it is eight times as fast at six times the cost. Her $97 demo is the proof. Sol sits at the other end: cheaper and fast enough for most engineering tasks.

Practical Application

Try a Dot and send feedback

If you are on Pro, Business, or Enterprise, your Dot is already in ChatGPT and the Codex app. Give it a real errand, such as shopping or scheduling, and watch for where the thread types (Dot, chat, Codex) confuse you. Claire asked for feedback on this very point.

Audit your team's vibe-coded tools

List the internal dashboards, prototypes, and reports your team has built with AI. For each, ask who should see which data. If any depend on one person's credentials, test whether Sites with connectors would let each viewer see only what they are allowed to see.

Start a shared Space

Create a Space with your agent as a running scratch pad for to-dos, decisions, and notes, and @mention ChatGPT to update it. It is a low-risk way to test whether human and agent collaboration in one doc works for your team.

Use fast vision for repetitive sorting

If you do a task that means scrubbing through many images or frames, such as picking video thumbnails, try a fast model with vision once the Decisions API is released. Claire went through 100 frames in under 10 seconds.

Questions to Consider

  • If your team's source of truth for docs and slides moved from Google Drive or Slack to AI-native Spaces that agents can edit, what would you gain and lose?
  • Which of your team's vibe-coded dashboards pull data that only some viewers should be allowed to see, and how are you enforcing that today?
  • If a frontier model responded eight times faster, what product in your roadmap would feel different to the user, and is it worth six times the cost?
  • Do you want one general agent handling all your work, or several small agents each hired for one job, and why?

Bottom Line

Claire's short list from DevDay: try Dots and send feedback, look at Spaces and Sites if your team struggles to collaborate on AI-made work, and watch the Decisions API with vision and ultra fast mode because they hint at AI products that were not possible before.

Resources Mentioned

ResourceTypeWhy it was mentioned
OpenAI DevDay 2026EventThe keynote and sessions where all of these releases were announced.
ChatPRDProductClaire's own product, built early on the idea that people and agents want to collaborate on the same document.

Tools & Products

Tool / ProductWhat it doesWhy it was mentioned
DotsAlways-on agent across ChatGPT, Codex, Slack, and its own computerThe headline release that Claire tested early.
ChatGPT SpacesShared workspace for docs, slides, and artifacts with humans and agentsCalled the most underhyped announcement.
ChatGPT SitesBuilt-in app builder with sharing and connectorsLets teams share vibe-coded tools with the right data permissions.
GPT-6.1 SolFast, low-cost modelHer workhorse for most engineering tasks.
Decisions APIFast constrained answers, now with visionUsed to pick podcast thumbnails and label a hot dog.
GPT-6 Astra ultra fastFrontier model at 8x speedPowered the sketch pad and 3D game demos.
CodexOpenAI's coding agentWhere Dots and ultra fast are available.

Experiments to Try

Experiment: Fast vision for sorting

  • Hypothesis: a fast vision model can pick the best frames from a long video faster and as well as a person.
  • Change: run frames from one video through the Decisions API with a fixed set of answers, such as "good" or "awkward".
  • Test: compare its picks against your own picks for the same video.
  • Measure: time spent and how many of its picks you would have chosen yourself.

Experiment: Real-time prompt-driven prototype

  • Hypothesis: a very fast model makes a prototype that changes with each prompt feel more usable than a slow one.
  • Change: build a small app where a prompt redraws part of the screen, as Claire did with a 3D game.
  • Test: run it at normal and ultra fast mode with a few users.
  • Measure: response time felt by users, how long they play, and the cost per session.
AI PM course

Everyone hears the same episodes.
Few can do what they describe.

Start for free