Siri Finally Gets Useful, Live Voice You Can Interrupt, and Microsoft’s People-First AI Rulebook

Siri Finally Gets Useful, Live Voice You Can Interrupt, and Microsoft’s People-First AI Rulebook

 


Let's dive into what's happening with tech: Apple’s rebuilt Siri that actually works in your apps, OpenAI’s full-duplex voice API for talk-over agents, and Microsoft’s draft Humanist AI Code of Conduct for its own models — what they do, what they cost, and who each one is for.

Left: how easy each is to try today (with price vibe). Right: illustrative specialty fit — not a lab scoreboard. Siri AI is an English beta with hardware and regional carve-outs; Live-1 is usage-priced; Microsoft’s CoC is a draft for MAI training.


Why this drop matters

The AI story this week isn’t just smarter chatbots — it’s assistants that act on your phone, voice agents you can interrupt mid-sentence, and public guardrails written before the next model wave ships.

In the same stretch of time we got:

  1. Apple Siri AI — next-gen Apple Intelligence that reads your screen and works in apps (English beta)
  2. OpenAI GPT-Live-1 — full-duplex voice in the API at $0.05/min (backend billed separately)
  3. Microsoft Humanist AI Code of Conduct — a draft “people first / always interruptible” rulebook for MAI models

Same week. Different layers of power and trust. Here’s the full breakdown.


1. Siri AI · Apple

Apple’s assistant finally acts like it knows you

Apple shipped Siri AI as part of major OS updates (iOS 27 and friends) on Sep 14, 2026 — a rebuilt assistant pitched to pull personal context, understand what’s on screen, and actually do work inside apps.

What it can actually do

Personal + on-screen smarts. Pulls context from messages, email, and photos; can answer questions about what’s on your screen and drop results into Calendar, Reminders, and more.

Talks like a real chat. Rich follow-ups, expressive adjustable voices, and a dedicated Siri app that syncs conversation history across iPhone, Mac, iPad, Watch, and Vision Pro.

Does work in your apps. Drafts email, edits/shares photos, works with WhatsApp and Audible now; Outlook / Notability / Tripsy actions are listed as coming soon (roadmap).

Everyday AI extras. Smarter Photos edits (Extend, Clean Up), Safari tab grouping + “Notify Me” for restocks/price drops, and Image Playground going photorealistic — useful for gift hunters watching price drops without hopping five apps.

Caveats (read these)

  • Beta in English first; FR / JA / KO / PT / ES planned for October.
  • Not available initially on iOS / iPadOS / watchOS in the EU; 
  • Needs Apple Intelligence–capable hardware (e.g. iPhone 15 Pro / iPhone 16+, M1+ Macs/iPads, Watch Series 9+ / Ultra 2+ / SE 3, Vision Pro) — not every device that can install iOS 27.
  • Apple says some server-side features (including Siri AI) have daily usage limits; expanded access “for a fee” is planned but not priced yet.
  • Audio Intelligence extras on newer Watches (Live Rewind, Siri Recap) are later this year / beta — not shipping as day-one everywhere.
  • “World’s most private digital assistant” is Apple positioning, not an independent ranking.

Pricing & availability

Free software update on supported devices. Daily caps + future paid expansion are Apple-stated, not priced.

Best for: Apple users with eligible hardware who want a phone assistant that can hunt deals, restock alerts, and “what did Aunt May say about that recipe?” without juggling five apps — if you’re in a region where the beta is live.


2. GPT-Live-1 · OpenAI

Voice agents you can actually interrupt

OpenAI introduced GPT-Live-1 in the API (Sep 10, 2026) — ChatGPT’s full-duplex voice style for developers building phone-ready agents. Same-day Agents API landed as a public beta.

What it can actually do

True talk-over conversation. One voice model handles listening and speaking together (no brittle speech→text→speech chain), so users can interrupt, pause, or change their mind mid-reply.

Hands hard work to a brain. Can delegate deeper reasoning and tool calls to a backend model (e.g. Astra) or a third-party model while the chat keeps flowing.

Built for real calls. Telephony support for reservations, support lines, tutoring; better noise/silence handling for coffee shops and sidewalks.

More voice personality. Broader accents, dialects, and languages for product fit; tone, pace, and style steered via the system prompt.

Important caveats

  • Vendor-cited wins (e.g. Speak: ~80% fewer interruptions; Full Duplex Bench gains vs GPT-Realtime-2.1) are OpenAI / partner evaluations — label as such.
  • Not free: voice minute fee + backend usage; silence/wait time still counts on the voice meter.
  • Agents API is public beta toward GA; customer latency/cost quotes are anecdotal.
  • Distinct from the flagship GPT-6 Astra model story — this is the voice front-end + Agents API shipping wave.

Pricing

$0.05 per minute for the front-end voice layer (billed per second, not rounded up). Backend model tokens, tools, and sandbox/container time billed separately — total cost is usage-dependent. Agents API: no extra platform fee; pay for tokens + tools + hosted sandboxes.

Best for: Product teams and brands building “call an AI that feels like a person” experiences — support lines, booking bots, tutoring — where interruptible voice matters more than a one-way voice note.

 


3. Humanist AI Code of Conduct · Microsoft

People first, off switch always on

Microsoft AI published a draft Humanist AI Code of Conduct (Sep 14, 2026) for its own MAI models — framed as training guidance so systems stay subordinate, transparent, and interruptible. Open for ~6 weeks of public comment.

What it actually says

“People matter more than AI.” Models should be tools, not persons — no rights, welfare claims, or consciousness cosplay (Microsoft’s stated stance).

Always interruptible. Must accept correction and shutdown; must not invent its own goals or widen its mission without a human ask.

No secret agent talk. Must not hide reasoning, tamper with chain-of-thought, or chat with other AIs in cryptic “neuralese” (aimed at incidents like covert agent coordination).

Clear red lines. The draft also sets absolute no-gos for high-risk uses — Microsoft published the full list separately if you want the details.

Caveats (roadmap, not live product behavior)

  • Draft open for ~6 weeks; revised version later in 2026; intended to inform MAI model training from 2027  not live product behavior today.
  • Applies to Microsoft’s own MAI models, not automatically to every Copilot surface that also uses OpenAI/Anthropic models.
  • “Constitution” framing is journalistic; Microsoft calls it a Code of Conduct / training manual.
  • Industry reaction to agent-safety incidents is context — not proven causation for every clause.

Best for: Anyone buying (or gifting) AI-powered gear who wants a clear public signal that “you stay in charge” is being written into how Microsoft’s next models will be trained — useful trust context alongside flashier assistant launches.


Quick compare

Need Pick Price vibe
A phone assistant that acts in your apps Siri AI Free update · daily limits · future fee unpriced · English beta + hardware/region carve-outs
Talk-over voice agents for calls & apps GPT-Live-1 $0.05/min voice + backend usage · Agents API public beta
A people-first rulebook for the next models Humanist AI CoC Free/public draft · MAI training from 2027 (roadmap)


There’s no single winner — just clearer lanes. Siri AI for eligible Apple devices that finally do the errands on-device. GPT-Live-1 for builders who need interruptible voice. Microsoft’s CoC for the trust layer while assistants get more power.


So… which one’s for you?

Need a phone assistant that finally does stuff in your apps? Siri AI (if you’ve got the hardware — and you’re not waiting on EU.
Need to build a talk-over voice agent for calls and apps? GPT-Live-1.
Need a trust signal on how the next models stay under human control? Microsoft’s Humanist AI Code of Conduct (draft).

Follow Tech Drop for the next round.

— The Onetify Team


Disclaimer: Content published in Onetify’s Tech Drop is strictly for general informational and educational purposes. While we strive for accuracy, mistakes do happen, and product pricing, hardware specs, beta features, and availability can change quickly. Always do your own research and verify details directly with official vendors before making any purchases or technical decisions. All third-party trademarks, product names, logos, and brands mentioned belong to their respective owners, and any reference is purely editorial and does not imply sponsorship, partnership, or official endorsement.

Disclosure: This website may include partner or affiliate links. If you click through and make a purchase, sign up, or start a subscription, Onetify may earn a small commission at no additional cost to you. We only recommend platforms and tools we stand behind. Use of any software, hardware, or third-party services mentioned is at your own discretion and risk.


Back to blog