CHAPTERS
- 0:00 – 0:30
GPT-6 Astra: what it is, why it feels like a step-change
Claire shares early-access impressions of GPT-6 Astra and why it’s made her feel “more ambitious” compared to prior models. She frames the episode around what OpenAI claims Astra is best at and previews hands-on demos across computer use, coding, and building real artifacts.
- •Astra one-shots tasks that stumped prior models (Fable, GPT‑5.6 Sol)
- •OpenAI positioning: most intelligent + aligned; strong across science, math, coding, knowledge work
- •Episode roadmap: computer use demos, coding breakthroughs, hardware/creative builds
- 0:30 – 3:02
Benchmarks, positioning, and ROI story: computer use as the killer feature
She quickly walks through the blog-post highlights and performance claims, emphasizing “computer use” across real software. Claire ties this to an emerging ROI narrative: models justify token spend by directly saving time on common workflows.
- •Computer use beyond browsing: Excel, Unity, Power BI, Blender, Jupyter, doc editors
- •“Work artifacts” tuned to style/brand/voice; better autonomy and steering
- •Time-saving examples (e.g., apartment hunting, thumbnails) as token ROI framing
- •Mentions benchmark deltas: Automation Bench and coding improvements; big jump in frontier math
- 3:02 – 3:32
Availability and pricing: who gets Astra first and what it costs
Claire outlines rollout timing and access tiers, then calls out pricing and speed modes. This sets expectations for when viewers can try Astra and in which environments it will show up.
- •Rollout: Daybreak enterprise first, then Plus/Pro/Enterprise
- •Access via API and AWS as it expands
- •Pricing: $10/M input tokens, $50/M output tokens
- •Fast mode expected to be notably quicker
- 3:32 – 8:06
Computer use demo #1: editing a complex node-based CRM workflow hands-free
Claire demonstrates Astra/Codex taking over Chrome to modify a node-based workflow in her CRM (Adio) to route inbound leads and draft personalized outbound emails. She highlights that Astra handles complex UI interactions—creating nodes, reconnecting logic, and shaping prompt outputs—without manual dragging and clicking.
- •Business context: routing new leads between Claire and Zach based on lead type
- •Goal: add an AI step that generates a customized email draft (subject/body) + scheduling CTA
- •Astra manipulates a complex node-based UI: adds objects, nodes, and connections autonomously
- •Takeaway: UI-heavy SaaS becomes more usable when AI can operate the interface
- 8:06 – 9:07
“UI is back”: why button-clicking agents change product thinking
She zooms out from the CRM example into a product thesis: earlier narratives pushed “no UI” (CLIs/MCPs), but strong computer-use agents make traditional UI valuable again. Claire argues this could be a lifeline for SaaS because humans like buttons—AI can press them.
- •Shift from “no UI” to “UI’s back” as agents reliably operate interfaces
- •Less need to force CLIs/MCPs if a UI-based product is the best expression
- •Computer use reframes how people interact with websites and software day-to-day
- 9:07 – 12:39
Computer use demo #2: Flora thumbnail asset generation from photos
Claire tests Astra in Flora, a node-based creative tool, to generate podcast thumbnail assets using newly captured images. She contrasts human toil (dragging nodes, prompting repeatedly) with Astra’s ability to inspect existing workflows, import images, select models/aspect ratios, and run generations automatically.
- •Prompt: generate designer-ready thumbnail assets using existing Flora workflows
- •Astra navigates messy node graphs, imports images, selects gpt-image-2, and generates outputs
- •Human role shifts to taste/QA (avoid uncanny faces, extra fingers)
- •Practical impact: reduces repetitive creative production overhead
- 12:39 – 15:10
Browser use for QA: automated testing with console inspection and race conditions
Claire describes using Astra to QA a preview branch of ChatPRD by clicking through flows, refreshing, and monitoring browser console errors. She emphasizes how agents can run long, tedious test sessions and surface edge cases that are hard for humans to reproduce consistently.
- •Astra tests a deployed preview branch by interacting with the UI like a user
- •Inspects console logs for errors and validates streaming/persistence edge cases
- •Runs extended QA sessions (~1h45m) to find intermittent race conditions
- •Recommendation: developers should use computer-use agents as QA assistants
- 15:10 – 18:43
Coding breakthrough: ChatPRD “product intelligence” feature finally clicks
Claire explains a long-running engineering challenge: ingesting messy product data, deduplicating insights, generating a credible product wiki, and verifying it against sources. Astra delivers a near-complete solution in one pass (plus minor follow-ups), achieving insight quality she couldn’t get from prior models.
- •Feature scope: ingest from Intercom/Granola/Linear/GitHub; store durably; process + dedupe
- •Outputs: top priorities, trends over time, opportunities, and an auto-generated product wiki
- •Astra delivers high-quality insights and a working implementation with minimal iteration
- •Highlights verification against external sources and exposing the wiki via MCP
- 18:43 – 22:15
Hardware hacking: Divoom MiniToo reverse engineering, CLI control, and live streaming text
Claire revisits her “Everest” device: a locked-down Bluetooth Divoom MiniToo with no official API. Astra helps build tooling to control it live (draw with mouse, type text) and a CLI that can fetch info (latest episode) and stream it to the display—including animation.
- •Context: proprietary device previously very hard to hack; now reliably controllable
- •Astra enables live app control + a CLI interface for scripted interactions
- •Demo: fetch latest podcast episode and stream a formatted blurb to the device
- •Extends to notifications: integrates with Codex hooks to signal when attention is needed
- 22:15 – 24:17
Desktop app build: a 90s AIM-style Mac client for Codex threads
Claire shows an Astra-built Mac app that wraps her Codex threads in an AOL Instant Messenger-style UI. Threads become “buddies,” complete with away message and active-typing states, illustrating how Astra makes novelty UIs and personal tooling feasible in a single shot.
- •AIM nostalgia interface mapped onto live Codex conversations
- •Threads appear as contacts; quick switching between projects and chats
- •Functional touches: screen name, away message, active writer/lockout behavior
- •Theme: one-shot custom desktop software is now practical
- 24:17 – 27:18
Blender + 3D generation: Barbie Fashion Designer Bench
Claire introduces her personal benchmark: recreating Barbie Fashion Designer vibes using Blender-generated assets and a customizable character. Astra produces a functional prototype with clothing/hair/skin toggles and a runway walk—imperfect, but a clear leap over earlier attempts.
- •Barbie Bench as a playful but revealing 3D asset+interaction test
- •Outputs: character model, wardrobe/accessory toggles, hair and skin tone variations
- •Runway walk animation demonstrates end-to-end 3D workflow progress
- •Acknowledges limitations: human characters/shoes/pants still rough
- 27:18 – 28:49
A 3D “family app” world for kids: walking to places, logging habits, SEL check-ins
She shares a more polished 3D example: a navigable world for her children to record reading time, collect rewards, and do emotions/weather-style SEL activities. Claire stresses she didn’t code it—Astra generated the experience from a simple prompt, expanding what personal software can be.
- •Goal: gamify homework/reading with a 3D journey and interactive locations
- •Examples: walk to library to log reading; “Moon Water Pond” for emotions/SEL
- •Keyboard-controlled exploration; suggests agents could also operate it
- •Takeaway: rapid creation of personal, meaningful 3D experiences
- 28:49 – 29:19
Figma finally works with computer use: auto-building YouTube thumbnails
During a time crunch, Claire asks Astra to use Figma to assemble thumbnails for the episode. She notes Figma control has historically been difficult for computer-use agents, making this a notable improvement.
- •Real constraint: designer unavailable; needs thumbnails quickly
- •Astra operates Figma to compose a classic YouTube thumbnail layout
- •Signals better reliability in complex design tools beyond browsers
- 29:19 – 32:19
Wrap-up: what to try first with Astra (and what’s next)
Claire summarizes Astra’s strengths—especially computer/browser use and higher ambition coding—along with small practical notes on speed and “not being annoying.” She previews follow-up comparisons and invites viewers to share what they build once access rolls out broadly.
- •Best first use: computer use for complex UI (CRMs, node tools, forms, design tools)
- •Coding impact: more ambitious builds (insights pipelines, hardware hacks, desktop apps, 3D)
- •Personal notes: slower but acceptable; strong daily-driver potential
- •Next: blind taste test/bench comparisons vs other models; call to action to build and share
