CHAPTERS
- 0:00 – 1:00
DevDay recap setup: Codex reset button and a flood of releases
Claire opens with her on-the-ground DevDay recap from San Francisco, framing the day as exciting but overwhelming due to the sheer volume of announcements. She sets expectations: highlights first, then early opinions on a few features she tested.
- •Attended keynote and sessions in person; “Codex reset button” moment
- •DevDay theme: rapid-fire announcements across agents, models, and developer platform
- •Goal: help viewers prioritize what matters to them
- •Will share early hands-on impressions (not final reviews)
- 1:00 – 2:01
What Dots are: always-on long-running agents across ChatGPT, Codex, phone, and Slack
Claire introduces Dots as the headline launch: persistent agents that can operate across multiple OpenAI surfaces and channels. She outlines their core promise—an “always-on delegate” that can work in the cloud, locally, and in team contexts.
- •Dots positioned as long-running, cross-surface agents
- •Work across ChatGPT and Codex; can run on their own computer and access yours
- •Multi-channel access: text, “call your Dot,” and Slack integration
- •Framed as a personal delegate that can initiate work and threads
- 2:01 – 4:32
Hands-on with Dots: single agent, dedicated computer, and multi-channel UX questions
She shares early access observations: currently users get one Dot, with its own virtual machine and apps. While the capabilities look strong, she finds the interaction model (calling, transcribing, surfaces) and product boundaries still unclear.
- •Currently issued a single Dot (for Pro/Business/Enterprise and certain plans)
- •Dot has an avatar and a dedicated computer with browser and developer-relevant apps
- •Dot can also be granted access to the user’s local computer
- •UX uncertainty: how calling/voice/transcription fits together
- •Information architecture confusion across Dot vs chat vs Codex threads
- 4:32 – 7:05
Dots performance vs competitors: strong “bones,” rough edges, and where it fits
Claire describes Dots doing real work well—shopping, coding, proactive reminders—while still feeling early in product polish and positioning. She contrasts Dots with GroqBots (micro-agents) and Muse (consumer-friendly UX), and delays a definitive verdict.
- •Dot work quality is strong: shopping, coding, proactive schedule help
- •Proactivity and memory feel promising
- •Rough edges: surface area and thread model still confusing
- •Personal preference: likes hiring micro-agents (GroqBots) vs one agent to rule them all
- •Muse feels more approachable; full Dot review planned after maturation
- 7:05 – 8:06
Spaces: AI-native collaboration on docs and slides with humans + agents
Claire calls ChatGPT Spaces an under-hyped enterprise feature: a collaborative workspace for documents, pages, and slides designed for co-working with agents. She highlights why this matters for businesses: sharing, permissions, and agent-native workflows.
- •Spaces enable collaboration on documents/pages/slides with humans and agents
- •Enterprise appeal: sharing controls, permissions, governance
- •Solves “creation is easy, editing/collaboration is hard” problem for slides
- •Parallels her earlier work building agent-human collaboration docs
- •Watches for real enterprise adoption vs existing sources of truth
- 8:06 – 10:38
A practical Spaces workflow: the “personal scratchpad” and stream-of-consciousness to-dos
She shows a concrete use case: her Dot creates a personal scratchpad space for ongoing to-dos and notes. Claire relates it to her own “Claire’s Notebook” workflow—lightweight, bullet-point planning shared across multiple agents.
- •Dot-generated “personal scratchpad” space as a daily operating hub
- •Claire’s preference: loose bullet lists over heavy Notion-style structure
- •Uses collaborative docs for to-dos, decisions, talk prep, and recording plans
- •Ability to @mention and update docs with ChatGPT/agents
- •Envisions the same pattern extending to slides and other artifacts
- 10:38 – 13:10
Sites + connectors/plugins: sharing internal AI tools with real data permissions
Claire explains how Sites (OpenAI’s deployable app builder) is being integrated with Spaces and upgraded to bundle connectors/plugins. She emphasizes enterprise impact: share a dashboard or tool powered by Snowflake (or other sources) while enforcing per-user access controls.
- •Sites: OpenAI’s web app builder that’s easy to deploy and share
- •New upgrade: bundle connectors/plugins into a Site
- •Example: shared dashboards pulling data via Snowflake plugin/connector
- •Per-user auth and permissions determine what teammates can see
- •Positions this as a solution to the “what do we do with vibe-coded internal tools?” problem
- 13:10 – 14:41
Model and platform highlights: GPT-6.1 Sol for cheaper, fast workhorse performance
Shifting to what she cares about most—models and infra—Claire covers GPT-6.1 Sol as a fast, cost-effective option. She compares it to Astra: slightly less premium intelligence but dramatically cheaper, making it ideal for many engineering tasks.
- •Dev feedback loop: everyone wants models “faster and cheaper”
- •GPT-6.1 Sol positioned as near-Astra capability at Sol speed and low cost
- •Pricing comparison: Sol far cheaper than Astra (as presented in the talk)
- •Claire’s usage: Sol as a dependable mid-line model for engineering tasks
- •Astra still her favorite when she’s willing to pay for top performance
- 14:41 – 15:12
Decisions API with vision: ultra-fast constrained outputs for real apps
Claire introduces OpenAI’s Decisions API as a “Jev competitor”: constrained-answer, low-latency decisions with Luna-like intelligence. The standout is vision support, enabling quick classification and selection tasks that are hard to do with non-vision decision models.
- •Decisions API: predefined answer set + very fast responses
- •Positioned as a Jev-style component in AI architecture
- •Key differentiator: includes vision (“computer eyeballs”)
- •Enables fast, reliable decisioning for product workflows
- •Not yet broadly released; she plans head-to-head evaluation later
- 15:12 – 16:12
Demo: picking better podcast thumbnails by scanning frames for non-awkward faces
She demonstrates a real creator workflow: automatically scanning ~100 video frames and selecting flattering options for thumbnails. The Decisions API performs this quickly, solving a recurring pain point where many models are slow or unreliable.
- •Thumbnail creation pain: scrubbing video to avoid awkward frames
- •Built a tool using Decisions API vision to evaluate many frames rapidly
- •Example result: multiple “good face” candidate frames returned quickly
- •Highlights practical value over novelty: saves real production time
- •Contrasts with Jev limitation: no vision in Jev
- 16:12 – 17:14
Demo: “Hot dog / not hot dog” and why low-latency vision decisions matter
Claire adds a playful classification example to show speed and confidence: identifying a hot dog in under a second. The point is broader—fast vision-based decisions unlock UI and product patterns that feel instant.
- •Binary vision classification demonstrated with high confidence and low latency
- •Reinforces the utility of constrained outputs for production apps
- •Pairs well with event-driven interfaces and rapid pipelines
- •Positions as an architectural missing piece alongside generative models
- •Sets up excitement for broader Decisions API availability
- 17:14 – 18:46
Astra ultrafast: 8x speed, 6x cost—latency as a new product primitive
Claire dives into Astra ultrafast as the most “future-feeling” release: frontier intelligence delivered at much lower latency. It’s expensive, but she argues the product implications are huge—enabling near-real-time interactive experiences not previously feasible.
- •Modes: normal, fast (2x), ultrafast (8x) in Codex/API contexts
- •Cost tradeoff: ultrafast is significantly more expensive (6x)
- •Potential shift: high intelligence at near real-time speeds enables new apps
- •Latency still a few seconds—close, but not perfect real-time yet
- •Frames ultrafast as a catalyst for new UX patterns and interactive systems
- 18:46 – 19:46
Ultrafast demo #1 — The Other Pencil: real-time collaborative SVG drawing
She shows an AI-assisted sketchpad where Astra “draws alongside you,” generating SVG strokes that match the user’s doodles almost immediately. The demo illustrates how responsiveness changes the feel of co-creation from batch to interactive.
- •Human draws simple shapes; Astra responds by generating matching SVG art
- •Astra’s strength: quick, high-quality SVG generation
- •Interaction feels “almost too fast,” enabling playful co-creation
- •Demonstrates a new category: real-time creative copilots
- •Highlights the experiential leap driven by lower latency
- 19:46 – 21:18
Ultrafast demo #2 — Little Starship: prompt-driven 3D world edits + physics changes
Claire presents a 3D scene with characters that can be modified live via prompts—adding rooms, objects, and even changing physics like gravity. Astra ultrafast powers rapid scene re-rendering, hinting at dynamic games and interactive worlds built on-the-fly.
- •3D environment with characters; world changes created from prompts
- •Examples: build a bathroom, switch to a tropical island, remove gravity
- •Ultrafast enables near-live rendering of new 3D objects and scene logic
- •Maintains a log of changes to the world/ship as state evolves
- •Shows how kids engaged with iterative prompting as gameplay
- 21:18 – 24:20
Cost reality + final navigation advice: what to try first and what else launched
She notes the “$97 for 30 minutes” cost of the ultrafast demo and reiterates that it’s currently pricey experimentation. She closes by listing other DevDay announcements (Agents API, computer use, plugins, plan changes) and her top recommendations: try Dots, evaluate Spaces/Sites for collaboration, and watch Decisions API + ultrafast.
- •Ultrafast demo was expensive in practice; highlights current pricing constraints
- •Other launches mentioned: Agents API, computer use, Codex primitives, plugin updates/monetization rumors, new plans (Pro 500)
- •Her suggested focus: test Dots, especially for proactive productivity
- •For companies: look hard at Spaces + Sites for governed collaboration/sharing
- •For developers: pay attention to Decisions API with vision and ultrafast’s product potential
