OpenAI Release Bonanza; Anthropic Plays Safe; 'Adult' AI Coming

ChatGPT Becomes App Platform

What’s new: You can now use third-party apps inside ChatGPT. Early integrations include Canva (generate designs and pitch decks), Zillow (interactive home search), Spotify (create playlists) and more - all invoked in plain language without leaving the conversation. OpenAI’s new Apps SDK handles the plumbing, and monetization is on the roadmap.

Why it matters: Instead of opening a bunch of tabs, you’ll talk to one interface -ChatGPT - and it will “tap” the right app for you. Nick Turley, head of ChatGPT, said the next six months will push ChatGPT to feel “a little more like an operating system.”

🧩 Jargon buster: SDK (Software Development Kit) = the starter kit developers use to build compatible apps.

AgentKit and the Visual Agent Builder

What’s new: OpenAI’s AgentKit gives developers a visual canvas to design agents (logic, tools, file search, approvals), embed a customizable chat interface on websites and run evaluations to test how well an agent behaves. Partners like Box, Canva and Evernote are already building with it.

Why it matters: This lowers the bar from “AI demo” to production workflows - think support triage, back-office automations, or sales assistants that actually take actions, not just chat.

🧩 Jargon buster: Agent = an AI that can choose tools and take steps toward a goal, not just reply with text

Codex Graduates; Deeper Workplace Hooks

What’s new: GPT-5 Codex - OpenAI’s coding assistant - is out of research preview with a Slack integration and more enterprise controls. Sam Altman said, “Almost all new code written at OpenAI today is written by Codex users,” and named Cisco, Duolingo, and Instacart as adopters.

Why it matters: The assistant now lives where developers already work - editors, terminals, and chat - so it can read context (files, diffs, tickets) and help ship real features faster.

🧩 Jargon buster: Enterprise controls = admin settings for privacy, permissions, and auditing.

Sora 2 Opens to Developers

What’s new: Sora 2, OpenAI’s next-gen video generator with stronger physics and synchronized audio, now has API access, allowing companies to generate short clips directly inside their own apps. Mattel is already using it to visualize product ideas more quickly.

Why it matters: Expect a wave of in-app video - marketing previews, product explainers, and storyboards - no film crew required. (Rights and consent tools remain a bit of an open issue.)

🧩 Jargon buster: API = a doorway that lets other apps programmatically use a service.

Mature Apps Coming Later (W/ Age Gates)

What’s new: Developers must currently ship 13–17-safe apps, but OpenAI says it will allow 18+ “mature” apps once age verification and controls are implemented.

Why it matters: Content boundaries are moving from vague promises to policy and enforcement, which is crucial for creators building health, finance, or adult-only experiences.

🧩 Jargon buster: Age gate = a verification system that restricts access to adult content or capabilities.

In Brief

• Apps you can try first: Canva for posters → pitch decks; Zillow for maps inside ChatGPT; Spotify for playlists - all inside one chat thread.
• Browser rumors: Dev Day chatter hinted at a possible OpenAI browser project, but nothing official yet.
• Hardware watch: The Jony Ive consumer device (always-on, voice-first) was mentioned again, but no launch date was given.
• Pulse and beyond: Expect faster iteration on ChatGPT features like Pulse as “apps inside chat” expand.

Read More: The Rundown AI, The Verge, OpenAI, Canva, Zillow, Spotify

OpenAI Locks In Compute with AMD (Update)

OpenAI announced a multi-year AMD partnership securing up to 6GW of compute (starting with 1GW of MI450s in 2026) and an option for up to 10% equity tied to deployment milestones - on top of a larger Nvidia commitment.

Why it matters: training and serving frontier models is now an energy-and-silicon sport. Pre-buying chips and power years ahead reduces risk, diversifies vendors and pressures rivals who don’t have this “compute bank.”

🧩 Jargon buster: GW (gigawatt) = power capacity. Big AI models need lots of it to run data centers at scale. Consider this: A typical U.S. nuclear power plant produces around 1 GW of electricity.

Read more: The Rundown AI, AMD, OpenAI

Anthropic Open-Sources Petri to Stress-Test AI Safety

Anthropic released Petri, an open system that uses agents to probe other models through thousands of simulated, multi-turn scenarios (fake companies, tools, goals).

A “judge” agent scores transcripts for issues like deception, leaks, or policy evasion. In tests across major models, Claude Sonnet 4.5 and GPT-5 showed stronger safety; others showed higher deception rates.

Why it matters: as agents get more capable, automated audits become essential - manual red-teaming can’t keep up with long-horizon behaviors. Petri gives labs and enterprises a way to continuously detect regressions before deployment.

🧩 Jargon buster: Red-teaming = deliberately trying to break or exploit a system to reveal weaknesses.

Read more: The Rundown AI, Anthropic

Builder Corner: Customer Support Agent In Minutes

OpenAI’s Agent Builder now lets non-experts wire up practical flows: route product vs. billing questions, ground answers in your docs, and ship an embeddable widget via ChatKit. The pattern also works for lead-qual, help desk, or triage.

Why it matters: this is the bridge from “demo” to “daily operations.” Clear routing + retrieval-grounded answers prevent hallucinations and speed real resolutions.

🧩 Jargon buster: Grounding = forcing answers to cite your approved documents so outputs stay accurate.

Read more: The Rundown AI, OpenAI

⚡ Quick Hits

• CodeMender: Google demoed an agent that finds and fixes security bugs automatically.
• ChatGPT scale: 800M weekly users; API traffic 6B tokens/min, per Sam Altman.
• Voice workflows: ElevenLabs shipped Agent Workflows for branching, stateful calls.
• Holiday commerce: Adobe forecasts +520% YoY surge in AI-assisted shopping traffic.
• Sora 2 (API Access): OpenAI’s latest video model now exposed to developers: better physics/continuity, multi-shot prompts, synchronized audio. Early use cases: ads, explainers, product demos that don’t need a film crew.

Today’s Sources: The Rundown AI, Adobe, AMD, Anthropic, ElevenLabs, Google, OpenAI The Verge, OpenAI, Canva, Zillow, Spotify