OpenAI COOKED

Tech/Product Highlights from “OpenAI Dev Day” (per the subtitles)

Mod retro “dev kit”

• A customizable, build-your-own-games device/workflow (“mod retro”).
• Mentions code/agents (“with codecs”) as part of the ecosystem.

“DOTS” — OpenAI’s always-on personal AI assistant

• Positioned as OpenAI’s answer to competitors like Grokbot, Muse, and “Instinct.”
• Core concept: always-on + proactive
  • Unlike typical chat that only responds to prompts, DOTS is described as thinking ahead and attempting to do what you’ll need before you need it.
• Integration channels:
  • Can connect through ChatGPT, Slack, and Teams.
• Agent architecture:
  • DOTS is described as its own agent/set of agents that runs in the cloud.
• Account/data access:
  • Can connect to Google/Gmail/Docs/Calendar and similar services to operate with your existing data.
• Design convergence:
  • Subtitles note a similar “soft geometric shape with eyeballs” design language across assistants (DOTS vs Grokbot), implying a broader market trend.
• Interaction with ChatGPT/Codeex:
  • DOTS “lives within the ChatGPT app” and can control ChatGPT and Codeex threads.
  • Review/opinion: the speaker says they don’t love this “inside ChatGPT” approach, but accepts the integration rationale (cloud environment + coding/control built on top of Codeex/ChatGPT).
• Usage/billing detail:
  • DOTS conversations don’t count toward ChatGPT usage.
  • But once DOTS controls ChatGPT/Codeex threads, that does count—important for cost expectations.

“Ultraast” / GPT-6 Astra via Cerebrus chips

• Presented as GPT-6 Astra powered by Cerebrus chips.
• Performance claim:
  • Up to 8× faster than Astra on Nvidia GPUs.
• Cost tradeoff:
  • About 6× more expensive.
  • Warning: tokens burn quickly
    • Example given: $1,000 in ~1.5 hours (early access).

New plan: Pro500 ($500/month)

• Higher usage limits:
  • 25× more than Plus (with the speaker reframing earlier tiers: a $200 plan was described as “20×,” now “10×”; $500 becomes “25×”).
• Takeaway:
  • Prices likely aren’t dropping proportionally at the frontier; faster/higher-quality compute costs more.

“GPT-6.1 Soul” — another frontier-leaning model

• Described as:
  • Affordable, fast, efficient, and close to “absolute frontier” quality.
• Pricing (quoted in subtitles):
  • GPT-6.1 Soul: $2/M input tokens and $10/M output tokens
  • Speaker also mentions “10 cents for a cash million input tokens” (likely an example/discount-style statement).
  • Compared vs GPT-6 Astra: $10/M input and $50/M output.
• Benchmark claims:
  • “GPT PDF” benchmark: behavior like Astra at a fraction of the price.
  • “OS World” (computer-use benchmark): performs very well; speaker suggests Astra may still win at the extreme top end, but cost advantage is major.
• Underlying idea:
  • OpenAI and Anthropic are both described as improving post-training so smaller, “inference-friendly” models approach frontier-model behavior (faster/cheaper).

Codeex Security Cloud

• A cloud service that continuously scans code for:
  • issues
  • vulnerabilities
  • ongoing alerts as code changes.

Codeex “cloud-native” direction

• Codeex is moving to the cloud; emphasis is on coding workflows that don’t require your laptop to stay constantly open.
• Bottleneck observation:
  • In Ultraast mode, inference feels very fast.
  • The slower portion becomes tool calling / running terminal commands, making local/tooling latency the limiting factor.

Tooling/Dev UX updates

• Refresh CLI controlled via voice (voice-based control).
• New code review experience in Codeex.
• Decisions API preview
  • Mentions a small, fast decision-making model (“lightest/smallest model”).
  • Speaker speculation: another model (called “Je(re)v”) is likely faster because it was built specifically for decisions rather than repurposed (contrasting with a model called “Luna”).

Plugins relaunch

• Subtitles say OpenAI is relaunching plugins after earlier criticism that “apps are dead” / plugins didn’t work well.
• New approach described:
  • Apps can be native inside the ChatGPT experience.
  • Single sign-in / token portability: log into third-party apps using your OpenAI account and “bring your tokens with you.”
  • Speaker frames this as especially generous and potentially unlike Anthropic’s approach.

“Notion” competitor / workspace: “o1/ChatGPT space”

• Described as a Notion-like workspace but fully native for team/agent workflows.
• Supports mixed content types (subtitles mention PowerPoint, spreadsheets, websites) plus agents.

---

Overall “Most Exciting” Takeaway (per the subtitles)

• While multiple updates sound exciting, the speaker specifically highlights Codeex cloud-native plus the improved assistant direction (DOTS: always-on/proactive) as some of the strongest themes.
• The transcript ends before naming a single definitive “top” announcement.

---

Main Speakers / Sources

• Main speaker/source: The video author/attendee (named only by context in the subtitles: attended Dev Day).
• Mentioned creators/companies/models (within the content):
  • OpenAI (Dev Day announcements)
  • Comparator assistants/models (examples mentioned): Grokbot, Muse, Anthropic (e.g., Opus 5.5, Sonnet 5.5)
  • Peter Steinberger / OpenClaw (referenced in the content)