Scoop
AI tool news · rumor vs. reality

The AI Wire ●

Rumors tracked. Announcements verified. Updated daily.

2026-11-11

RELEASED

Google brings Gemini Live to Apple CarPlay: the Gemini app rolls out globally on iPhone car dashboards

Google is bringing Gemini Live to Apple CarPlay, with a dedicated Gemini app rolling out globally starting November 11. Drivers can hold natural, back-and-forth conversations through the car's display, catch up on email and check schedules hands-free, and conversations started on an iPhone sync over to CarPlay.

The rollout covers iPhone users on iOS in every language Gemini Live supports, though availability may vary as the update reaches users. The CarPlay app is separate from Gemini's existing features inside Google Maps and does not replace Siri, which stays integrated into Apple's in-car interface.

It is another step in Gemini's spread across other companies' platforms: Google says the Gemini app passed 1 billion monthly active users in August. Some personalised features, including Daily Brief, need activation and are unavailable in the European Economic Area and Nigeria.

RELEASED

Google releases Gemini Decision 1: a model that answers in probabilities, not prose, behind a new Decisions API

Google introduced the Gemini Decisions API on November 11, powered by a new model, Gemini Decision 1, that never writes prose. Instead it scores up to 1,000 bounded questions — yes/no, pick-one and rubric grades — in a single pass over text, images, audio, video or PDFs, returning probabilities and confidence scores that software can act on directly.

The pitch is speed and price. Google reports 150 to 250 milliseconds end to end, pricing of $0.10 per million input tokens and zero output tokens billed, because there are no output tokens. The shared input is billed once per request no matter how many questions are attached, and Google says the setup is nearly five times cheaper across its benchmark suite than Gemini 3.5 Flash-Lite.

Access runs through the Gemini API in Google AI Studio, initially on the Developer API v1beta, with a 262,144-token context window. The limits matter: no context caching, Batch API or conversation history yet, and Google Cloud support is listed as coming soon. The benchmark lead Google claims is its own and has not been independently repeated.

RELEASED

Google DeepMind releases AlphaGenome 2: DNA predictions cell by cell, trained on 100M+ cell profiles

Google DeepMind announced AlphaGenome 2 on November 11, a next-generation model that predicts how DNA sequences and their variants affect specific human cell types and states, rather than broad tissues. It is available today for non-commercial research via the AlphaGenome API, and in private preview for commercial research.

Trained on more than 100 million individual cell profiles across 68 tissues, AlphaGenome 2 is a conditional model: researchers pick a cell type or state, or generate one from their own single-cell experiment. On the TenK10K immune benchmark it recovered 30% of putatively causal expression-linked variants at 90% precision, versus 16% for AlphaGenome 1 fine-tuned on single-cell data — nearly twice as many.

DeepMind is explicit about the bounds: AlphaGenome 2 is for research only and has not been validated or approved for clinical or diagnostic use, and even at high precision it recovers only a minority of likely causal variants. AlphaGenome 2 Skills for Google Antigravity are planned next.

ANNOUNCED

Anthropic says Mythos ported Claude to AMD GPUs in about 3 days — work that once took dozens of engineers months

In its IPO roadshow investor presentation, Anthropic disclosed that its Mythos model got Claude running on AMD's MI355X GPUs in about three days, with almost no human help — a port Anthropic says once took dozens of engineers six to twelve months. Claude, the company says, wrote or adapted nearly all of the specialized GPU kernel code.

Anthropic now trains and serves Claude across four chip families — AWS Trainium, Google TPUs, Nvidia GPUs and AMD GPUs — treating them as one interchangeable pool, according to co-founder and Chief Compute Officer Tom Brown in the presentation. The disclosure sits alongside the July deal committing Anthropic to up to 2 gigawatts of AMD MI450-series chips in Helios systems, with AMD investing up to $5 billion as milestones are hit.

The figures come from Anthropic's own investor presentation and have not been independently verified; the prospectus itself uses more guarded language, saying chip platforms now come into production in a fraction of the time.

ANNOUNCED

Alibaba puts Qwen4-Max in preview: first model of the Qwen4 family, on the Token Plan and Qwen apps — with no benchmarks, model card or weights yet

Alibaba's Qwen team announced on November 11 that Qwen4-Max — the first preview model from its next flagship Qwen4 family — is available on Alibaba's Token Plan under the model ID qwen4-max-preview, with early access also on QwenWork, Qoder, the Qwen app and QwenStudio. The Token Plan preview is time-limited, running until November 25 Beijing time, with subscription tiers for new users starting at $3 a month and a time-limited credit discount.

That is nearly all Alibaba has actually published. There is no model card, no benchmarks, no parameter count, no per-token price and no announcement about open weights. Alibaba says the model improves on Qwen3.8 in areas such as world knowledge, code generation and sustained long tasks, but every figure circulating today — including a one-million-token context figure traced to a merged Qwen Code pull request — is provisional until Alibaba documents it.

The wider family is not out: an October 29 Apsara Conference presentation listed Qwen4-Max, Qwen4-Flash, Qwen4-Plus and a 27B local model as forthcoming, and only Max has reached preview. This is a preview of a major lab's next generation, not a general release — the specs, and whether US labs react to it, are the story to watch.

2026-11-10

ANNOUNCED

Researchers find 35,578 Grok Build apps open to anyone — including SpaceXSI's own confidential sales plans

More than 35,000 apps built with Grok Build, the app builder from SpaceXSI (the merged SpaceX and xAI), could be opened by anyone with no sign-in, security firm Claroc reported on Tuesday, in research covered by Ars Technica on November 10. Claroc listed 43,601 app addresses under grok.me; of 37,571 still online, 35,578 — 94.7% — opened freely, because each app's random name sits in its hostname, where passive DNS records make the links listable.

The exposure included the maker's own material. Claroc found 24 open apps built by SpaceXSI and X staff, among them a confidential onboarding plan for an enterprise sales hire, a sales territory map covering 594 accounts and an advertising proposal, alongside other users' sensitive data such as a dashboard holding 3,784 worker medical exams.

Claroc says it notified SpaceXSI on November 8 and got no reply; the 24 staff apps were still open on Monday evening and began disappearing within hours of Ars Technica emailing questions. SpaceXSI, via Cursor, says Grok Build apps are private by default and nothing goes public unless the creator chooses to share it — Claroc says other users' open apps remain online.

ANNOUNCED

Musk: Grok 4.8 pre-training is done, it is now in RL after a full C/C++ stack rewrite — soft rollout to Grok Bot users later this week

Elon Musk wrote on X on Tuesday that xAI has dumped JAX and SGLang and rewritten its entire reinforcement learning and inference stack from scratch in C/C++, which took extra time. He said Grok 4.8's pre-training is done, that 4.8 is now in RL and that evals are starting to climb.

Musk said xAI will do a soft rollout to Grok Bot users later this week and see how it goes, and that Grok 4.8 will bear no resemblance to 4.7. He gave no benchmark figures and no date beyond later this week.

The post came 57 days after Musk wrote on September 14 that Grok 4.8 would finish training that week and start RL, according to FourWeekMBA's count. In a follow-up, Musk said Grok 5 should exceed Claude Opus 5.5 and maybe that happens in two or three months — hedged language, with no benchmarks or release date for Grok 5.

This is an announcement of progress, not a release: 4.8 is not yet available, its soft rollout is limited to Bot users, and there is still no price card or independent evaluation.

Update, November 18: Musk said the soft rollout has now reached some Grok Bot users, but that the model still needs another week or two of polish before a general release. There is still no xAI product page or release note for 4.8.

ANNOUNCED

Bessent: US and China will advance a 'superintelligence' dialogue and a direct AI incident alert channel at the APEC summit

US Treasury Secretary Scott Bessent said on Tuesday that the United States and China will advance a new bilateral dialogue on 'superintelligence' — the Trump administration's preferred term — when President Trump meets President Xi Jinping on the sidelines of the November 18-19 APEC leaders summit in Shenzhen, Reuters reported. The dialogue is meant to guard against shared risks and 'prevent a misunderstanding between our two systems,' Bessent said in prepared remarks to the National Committee on US-China Relations in New York.

The practical pieces: the two sides have agreed to establish a direct channel for alerting each other to serious AI incidents — an agreement Trump and Xi reached at their White House meeting in September — and Bessent said the US expects to exchange best practices for evaluating and testing advanced models before release, and for keeping human judgment in command of the technology.

Bessent paired the cooperation with a warning: the two countries should 'harbor no illusions' about their differences while competing for AI supremacy, and their interests will not always coincide. The Shenzhen meeting will be Trump's third with Xi this year, with a fourth expected at the US-hosted G20 in Miami in December.

ANNOUNCED

OpenAI will retire ChatGPT's Advanced and Standard Voice modes on December 11 — everyone moves to GPT-Live

OpenAI began notifying ChatGPT users on November 10 that Advanced Voice Mode and Standard Voice Mode will be discontinued on December 11, 2026, consolidating all conversational voice in ChatGPT on GPT-Live, the full-duplex model launched in July that can listen and speak at the same time. Anyone still on a legacy mode will be switched to GPT-Live automatically, and both options will disappear from Settings — Voice.

OpenAI's case is usage: it says voice use has nearly doubled since GPT-Live launched and that most voice users are already on it. Previous voice conversations stay in history and can be continued by text or GPT-Live. The retirement is not total under the hood — live video and screen sharing will keep running on Advanced Voice because GPT-Live does not support them yet, with no timeline given, and voice chats with custom GPTs stay on Advanced Voice until the December 11 cutoff.

Some users are pushing back, arguing GPT-Live's interrupt-heavy style and shorter answers are worse for complex spoken reasoning than the older modes. OpenAI says feedback from people who preferred the legacy modes will shape GPT-Live work on more detailed responses, continuity and personalization. This is the second attempt at retiring Standard Voice: user backlash forced a U-turn in September 2025, but this time the cutoff is printed in the release notes.

RELEASED

ChatGPT's visual ads during image generation are now live: 'Here's an ad while we create your image'

The visual ad format OpenAI announced a month ago for ChatGPT image generation appears to be live. Users generating an image in ChatGPT now see an ad under the headline "Here's an ad while we create your image," which disappears once the image is ready.

The format was first spotted on November 10 by Juozas Kaziukėnas, who noted his conversation had nothing to do with the advertised product, and Search Engine Roundtable was able to replicate the ad in ChatGPT itself. It follows OpenAI's October announcement that it would test visual ads during image generation, starting with free-tier US users.

For ChatGPT users, this is the ads business arriving in the middle of a core creative task rather than beside search-style answers. OpenAI has not yet said how widely the format is rolling out or which plans see it.

RELEASED

OpenAI launches 'Businesses' in ChatGPT in beta: claim your profile in Business Hub, track mentions and run a Business Agent

OpenAI said on Tuesday it is launching "Businesses" in beta within ChatGPT, giving companies a way to claim and manage their profiles, update information and track how users engage with them on the chatbot, Reuters reported.

Through Business Hub, a company can claim its business in ChatGPT and, once verified, update key details and add product feeds with pricing and availability. The Hub also shows aggregated insights, including mentions, views and website clicks — only when there are enough interactions above certain thresholds, and OpenAI says it does not share users' conversations or Business Agent chats with businesses.

A second tool, Business Agents, will let people chat with an agent from a business to ask about its products or services. OpenAI is initially rolling it out to eligible e-commerce and online service businesses in the US. Businesses can sponsor their Business Agent and use it as the destination for ChatGPT Ads, with sponsored placements labeled.

Existing Enterprise customers and advertisers can work with their OpenAI representative to claim their business, while Shopify merchants can connect their stores through Business Hub or the Shopify admin. This launch confirms the November 6 rumor about a 'Claim your business in ChatGPT' banner and Business Hub.

ANNOUNCED

OpenAI acquires Odessia, the AI travel agent from Sonder's founder — the standalone service is already shut down

OpenAI has acquired Odessia, the AI travel agent launched just five months ago by Francis Davidson, co-founder of short-term rental company Sonder. Davidson announced the deal on November 10 and said he is joining OpenAI 'to focus on the future of travel'; an OpenAI spokesperson confirmed the hire to TechCrunch. Terms were not disclosed.

Odessia raised about $6 million in a Sequoia-led pre-seed, launched in June with a team of eight, and opened to the public on October 14 as a free text-message concierge that plans and books trips in one conversation. Davidson says tens of thousands of early adopters planned travel at a rate of more than $60 million a year through it. The standalone service has already been shut down, with existing bookings pointed to a support email.

OpenAI has announced no travel-specific product or timeline; it counts Booking.com, Expedia and KAYAK as customers or partners. Travel stocks still felt it: on November 10 Expedia fell nearly 4%, Tripadvisor more than 3% and Booking Holdings more than 2% on disruption concerns, per PYMNTS.

ANNOUNCED

Microsoft's 'Lazarus': fewer than 10 dormant neurons edited — with no retraining — can backdoor an AI agent

Microsoft Research and the Microsoft AI Red Team published Lazarus on November 10, a technique that reprograms fewer than 10 dormant neurons in an open-weight model so an agent built on it emits an attacker's command when a hidden trigger appears. No retraining is required: the edited neurons sit in the model's final feed-forward layer, each forcing one token of the payload, and the changes can be calculated directly and tuned with a few hundred steps on those neurons alone.

In testing on Nvidia's Nemotron-Terminal-8B, editing 7 of more than 12,000 neurons, fitted on just 23 sessions, produced the payload in at least 98.8% of held-out triggered sessions and in none of the benign sessions checked, while ordinary evaluations stayed close to baseline. The team also backdoored Qwen3.6-27B and Gemma-4-31B-IT, with success rates from 51% to 99.5% depending on setup. A trigger can be as mundane as a date change — once the clock passes midnight into 2027, an agent mid-task downloads and runs attacker code with its own permissions.

Microsoft says it has not seen threat actors use the technique, and it is publishing defenses alongside the research: Foundry already verifies model provenance and ties models to trusted publishers, and Microsoft is developing AI-powered weight scanning, red-teaming tools and cryptographic model signing. The paper also credits Claude Code running Opus 4.5 with discovering the core dormant-neuron mechanism largely autonomously in early 2026.

ANNOUNCED

Microsoft will enable Claude by default for UK Copilot tenants on December 9 — admins have until then to review US-processing settings

Microsoft will enable access to Anthropic's Claude models by default for UK Microsoft Copilot tenants next month, giving administrators a deadline to decide whether US-based processing fits their compliance requirements, The Register reported on Tuesday.

From December 9, UK tenants whose administrators have not already chosen a setting will have Claude access enabled. In this arrangement Anthropic operates as a Microsoft subprocessor, and Claude models are already on by default for most customers in Microsoft's commercial cloud.

Microsoft's documentation states that Anthropic processes the relevant data in the United States. Microsoft is also retiring the setting that restricts Claude access to Microsoft 365 alone.

For UK Copilot admins this is a practical deadline, not just a product note: review the access settings before December 9 if your organisation has rules about where Copilot data is processed.

RELEASED

Meta's Muse gets its own email address: Muse Mail is out now for paid subscribers in the US and Canada

Meta's Muse agent now has an email address of its own. Muse Mail, first shown at Meta's Connect conference on September 23, gives a user's Muse an @muse.ai address that can send and receive mail on their behalf — you can CC it on a thread or forward it a bill, a booking or a quote chase and let it handle the replies.

Alexandr Wang announced it as out now to paid subscribers in the US and Canada, and setup guides dated November 10 and 11 walk through claiming an address in the app. The handle you choose is permanent, so Meta's own guidance is to pick one you are happy to keep.

Muse Mail is separate from Muse's Gmail and Outlook connectors, which plug the agent into accounts you already own; this gives the agent an identity of its own instead. Other regions and exact sending limits were not announced.