The one OpenAI announcement that can actually make you money...

Key technological/product concepts and what was announced

“Dots” personal agent (OpenAI Dev Day)

• What it is: A GPT-6 Astra–powered “personal assistant/agent” where “dots” act as a unit of capability.
• How it’s used: Users name their agent, and it takes actions on tasks.
  • Example described: “Alfred” removing an old inventory API before shutdown.
• Demo focus: agentic software engineering
  • Tracing dependencies
  • Updating integrations
  • Running tests
  • Opening multiple PRs (implying an automated workflow)

Live demo difficulty

• Holly Lee reportedly demonstrated how Dots works in a live context (as opposed to polished/pre-recorded “slop”).
• The framing emphasizes that real-time demos are still hard.

Subscription and pricing changes (several tiers)

Pro 500 plan

• Price: $500/month
• Usage: 25× usage vs usual ChatGPT usage
• Speed: Access to an “ultrafast” tier running Astra at ~300 tokens/sec
• API pricing quoted:
  • $60 per million input tokens
  • $300 per million output tokens (as stated in subtitles)

$200/month plan

• Returned/sold again, but with usage cut in half (described in a “castrated” tone).

Budget tier: “GPT 6.1 soul”

• Positioned as near Astra intelligence for a fraction of the price
• Quoted:
  • $2 per million in
  • $10 per million out

Pricing theme: performance/speed tiers plus usage rationing for higher tiers.

---

OpenAI “Decisions API” (structured choice model)

• Core idea: Instead of prompting a large model to generate text and then parsing JSON with retries, the system uses a small “decision” model (described as Luna).
• How it works:
  • You provide a fixed list of allowed answers
  • The model returns one choice quickly (few hundred milliseconds)
  • This avoids JSON parsing/retry loops because output is constrained by design
• Example use case: routing/categorizing messages for a dating app, including a notion of “blackmail usefulness.”
• Broader trend tie-in: decision/routing models for type-safe, structured outputs.

---

Competitor/adjacent ecosystem references

Type-safe AI / “Jev” references

• Mentions prior content on Type-Safe AI and an allegedly similar model from an OpenAI researcher (referred to as “Jev” in subtitles).

Productivity-app “clones” and bots

• Tools described as clones:
  • Pages (Google Docs clone)
  • Space (Notion clone)
• A Slack bot in a similar spirit to “Claude Tag”
• Mentions “infinite tokens” in that context.

---

“Sign in with ChatGPT” for developers

• What it does: Lets app users log in using their ChatGPT account.
• Claimed benefit: Developers can make users “burn their own tokens instead of yours.”
• Launch partners mentioned: Devin, Notion, Vercel, OpenClaw, etc.
• Implicit risk/claim: If an app becomes popular, OpenAI may clone/compete (framed as a concern in subtitles).

---

Owning model weights / on-prem or local decision routing

• Suggests that for decision models, developers may want to own weights to avoid platform constraints.

Sponsor: Fastly Labs “Gliner 2.5 Decide”

• Open-weight decision model
• Runs locally (not closed-weight)
• Claim: up to 8× faster
• From the Gliner family
  • Stated as open source since 2024
  • 50M+ downloads claimed

Practical examples mentioned:

• Integrating with “OpenCode” to extract repo file paths/function names and route requests
• Developers using the approach to run a browser agent much cheaper than “Jev”

New model mentioned: Fastly “Glide”

• Claims to beat “Jev” on:
  • Intent routing
  • Fact checking
  • Hallucination checks

---

Tutorial/guide/review framing elements

• The video positions itself as a “breakdown of what was missed” at OpenAI Dev Day.
• Includes:
  • Pricing breakdowns
  • A product capability breakdown (Dots, Decisions API, Sign-in with ChatGPT)
• Promotes a deployment guidance angle:
  • Use open-weight decision models to route locally and own weights.

---

Main speakers/sources (as implied by subtitles)

• Sam Altman (announcements/demos described)
• Zuck / Mark Zuckerberg (mentioned earlier via a Muse reference/comparison)
• Holly Lee (live demo of Dots)
• Peter Steinberger (creator of “ClaudeBot,” referenced as prior work)
• Fastly Labs (sponsor introducing Gliner 2.5 Decide / Glide)