AI News Today — November 2, 2026: Banks and Retailers Block Muse AI Agents, Apple Intelligence Reaches China With Alibaba's Qwen, and ChatGPT and Codex Merge Their Limits

The big picture

The agent era met its first organized resistance this weekend. CNBC reports that five of America's biggest banks and retailers have moved to block personal AI agents like Meta's Muse — the same week The Wall Street Journal reported Muse is already pressuring Amazon's shopping model. Around that fight, the calendar filled up with dated, practical changes: Apple Intelligence starts testing in China on November 4, Anthropic's Claude watermark reaches four older models on November 3, Gemini's Gems go read-only on November 17, and in December ChatGPT and Codex start sharing one allowance. No new flagship model launched today — instead, the rules around the ones you already use are being rewritten.

What happened today

Muse AI: JPMorgan, Amazon and three more block personal agents like Muse

What happened: CNBC reported on November 1 that JPMorgan Chase, Charles Schwab, Amazon, Kohl's and Neiman Marcus have moved to block outside personal AI agents such as Meta's Muse, just as consumers start sending them out to find better savings rates, shop for deals and cancel subscriptions. A JPMorgan spokeswoman told CNBC the bank does not let outside agents log in for retail customers — agents can read some data, such as balances, but cannot send money. Amazon said shopping apps should respect a store's choice to take part, and asked Meta to remove Amazon from Muse. An audit by Billions Network of the published rules of 51 large U.S. retailers found only 11 had taken a clear position on AI agents: five block them on purpose, while 37 block them mostly by accident, through old rules written to stop data scrapers — and none of the 51 gives a verified agent a way to finish a purchase. CNBC noted the audit read published rules only, which does not prove how each store behaves in practice. Walmart, Shopify and Best Buy have taken the opposite route and agreed to work with Muse. Accenture's global banking lead Michael Abbott told CNBC that companies blocking agents may struggle long-term and should build secure ways for authorized agents to interact instead, citing Accenture research that 65% of consumers are open to a GPT-like financial assistant and 71% would welcome one in their primary bank's app.

Who it matters to: Anyone building a workflow on a personal agent — and any business deciding whether agents are customers or trespassers. This is the DoorDash "several hundred agent orders" story from last week growing teeth: the demand is still small, but the access fight has started before the volume arrives, and most retailers are blocking agents by accident rather than by policy.

What you can do: If you use Muse or a similar agent for shopping or banking tasks, expect it to work at Walmart, Shopify stores and Best Buy, and to hit walls at Amazon and the big banks — check before you build a routine around it. If you sell online, audit your own bot rules this month: right now, the likeliest outcome is that you are blocking agents unintentionally, with no verified-agent path at all.

Learn more: CNBC's report on the agent blocks, summarized and the WSJ's Muse-vs-Amazon shopping story, via Seeking Alpha

Muse AI gets a body on paper: the Max humanoid details firm up

What happened: Yesterday's unconfirmed humanoid rumor got a real source. The Information reports — carried by Reuters and multiple follow-ups — that Meta built a humanoid prototype called Max with Shanghai-based robotics firm Fourier in about three months. Max stands about 1.5 meters tall, weighs about 48 kilograms (106 pounds), runs on Nvidia's Jetson Thor chip, and has five-fingered hands that can pick up small objects, built-in speakers, and a face display that mimics expressions. Meta's vice president of product for robotics, Li-Chen Miller, spent almost two months at Fourier's Shanghai factory, and Max was shown at an internal robotics event in late August. Work on Max 2 reportedly began in September, with a redesigned torso, improved hands, a projector, and extra sensors for stairs. CTO Andrew Bosworth has described humanoids internally as the "physical embodiment" of Muse. Crucially, this is still not a product: a Meta spokesperson told The Information the Fourier work is "focused on internal research and experimentation, and we have no plans to sell these units," and no launch date exists. The Chinese manufacturing partnership is also drawing scrutiny, since U.S. regulators have treated connected foreign-made robots as a security risk.

Who it matters to: Anyone tracking where Muse goes next. Meta is doing the opposite of the robotics startups: it started with an agent that already has mass distribution — more than 10 million Muse setups in the U.S. and Canada in its first six weeks, by Meta's own count — and is now prototyping the body. But a robot that can pick up a bottle is a long way from one that can run a home, and Meta itself is saying research, not retail.

What you can do: File this under direction, not deadlines. There is nothing to buy, preorder, or plan around — the useful signal is strategic: Meta's agent ambitions do not stop at your phone, which is one more reason the access fight in the story above matters.

Learn more: Notebookcheck on The Information's Max report

Apple Intelligence comes to China in December, with Alibaba's Qwen inside Siri

What happened: Apple announced on its China Newsroom on October 31 that Apple Intelligence will open for testing in mainland China on November 4 in the latest iOS developer beta, with a public release alongside iOS 27.3 in December — two years after the features first reached the U.S. The first wave covers about 18 features, including Live Translation, Writing Tools, Visual Intelligence, Clean Up and Genmoji, notification and Mail summaries, and Type to Siri. In selected experiences — Siri, Writing Tools and Visual Intelligence — requests can draw on Alibaba's Qwen models for text and image understanding and content generation, standing in where other markets can call on ChatGPT. The integration is opt-in, users sign in with a Qwen account or Sign in with Apple, and Alibaba confirmed that requests and responses sent through Apple Intelligence will not be used to train its models. The release is iPhone-only — iPhone 15 Pro and Pro Max, iPhone 16 and later, iPhone Air and iPhone Duo — with iPad and Mac support waiting on further approvals, and it explicitly does not include Siri AI or the other next-generation features announced at WWDC 2026. Alibaba's stock rose 2.91% in European trading on Monday morning, November 2, on the news.

Who it matters to: Developers and businesses with users in China, first of all: the same Siri and Writing Tools surfaces now route to a different model depending on device location and Apple Account region, so output quality and behavior will differ by market. For everyone else, it is the clearest example yet of AI features fragmenting by jurisdiction — one product, different engines, different feature sets.

What you can do: If your app or content reaches Chinese iPhone users, test Writing Tools and Visual Intelligence in the November 4 beta rather than after the December launch, in Simplified Chinese and on real photos and documents. If you don't ship to China, no action — but note the pattern when planning any "global" AI rollout.

Learn more: 9to5Mac on Apple's China announcement

ChatGPT and Codex will share one usage allowance in December

What happened: OpenAI's Help Center article "Understanding ChatGPT and Codex usage limits," updated October 31, says that starting in December, the usage allowance on Plus, Pro and Business plans "will be shared across ChatGPT and Codex" and "measured in tokens," with more advanced models and complex tasks using more of it. The change arrives as ChatGPT Work's agentic features move into the main ChatGPT experience — Work and Codex already share one allowance today, and December extends that pool to ChatGPT itself. One category is exempt: "Instant chat will not draw from your usage." Combined remaining usage and reset times will show in Settings → Usage, and at the limit OpenAI lists three options: wait for the reset, upgrade, or buy credits for eligible features. OpenAI says it expects most users "won't need to wait for their usage to reset more often than they do now" — a forecast, not a guarantee. It has not published per-plan token amounts or an exact December start date, so there is no official before-and-after comparison yet.

Who it matters to: Anyone who both chats and codes on a paid plan. From December, a heavy afternoon of Codex agent runs competes directly with your ChatGPT use, and vice versa — the two budgets you have today become one budget with one clock.

What you can do: Starting now, check Settings → Usage weekly and note your rough split between chat and Codex while the allowances are still separate. That baseline is the only way you'll be able to tell in December whether the merged pool actually shrank your effective capacity. Anything that can live in Instant chat stays outside the meter entirely.

Learn more: A walkthrough of OpenAI's updated usage-limits article and The Neuron's November 2 roundup

Anthropic extends Claude's invisible text watermark to four older models on November 3

What happened: Anthropic emailed API customers that starting Tuesday, November 3, it will apply its text watermark to the output of four older models — Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 — on the Claude Platform, with the Amazon Bedrock rollout possibly a few days behind, according to reporting by The Decoder and Mashable. It is the same mark the newest models, Opus 5.5 included, already carry: a version of Google DeepMind's SynthID-Text that steers word choice statistically rather than adding hidden characters, and Anthropic says internal tests found no effect on quality or readability. The retrofit covers models launched before the EU AI Act's August 2, 2026 transparency start, ahead of the Act's December compliance deadline. There is a real caveat: on Opus 4.6 and Sonnet 4.6, the watermark applies only at default generation settings — changing temperature, top_p or top_k can leave output unmarked, per Anthropic's email as reported by The Decoder. Anthropic also stresses the limits of detection: a detected mark means Claude was likely involved at some point, not that it wrote the whole text, and the detector itself remains in private preview for eligible organizations such as regulators, media, researchers and compliance-bound enterprises — there is no public checker for text.

Who it matters to: Developers and compliance teams on older Claude models — from tomorrow, output provenance changes under you, and on the 4.6 models your own sampling settings decide whether the mark is even there. It also matters to anyone tempted to treat watermark checks as proof about a student or a writer: Anthropic itself says the signal is a probability, not a verdict.

What you can do: If you rely on watermarking for your own compliance, keep temperature, top_p and top_k at defaults on Opus 4.6 and Sonnet 4.6, and don't assume Bedrock output is marked on day one. If you evaluate other people's text, treat a mark — or its absence — as one weak signal, never as evidence on its own.

Learn more: Coverage of the November 3 watermark expansion and its caveats

Gemini Spark: no new Spark announcement — the Gems-to-Skills migration is now the story, 15 days out

What happened: Our daily priority check found no new Gemini Spark announcement. Spark remains folded into the Gemini app as Task mode — the Chat | Task switch, event-triggered multistep tasks across apps and the web, up to 15 tasks at once, for Google AI Pro and Ultra subscribers with Keep Activity on. What changed this weekend is the migration underneath it. In a hands-on test published November 1, XDA Developers confirmed the timeline and the escape hatch: on personal accounts, Gems become read-only on November 17 and are permanently deleted on December 17; Google will auto-convert Gems to Skills, but each Skill has a 100 MB file limit and oversized attachments are dropped largest-first, with formats like legacy Office documents and archives not carrying over. The portable part: Skills use the same SKILL.md format as Claude — Gemini accepts a single SKILL.md file, a skill folder, or a zip upload through Customize → Skills, and the XDA tester moved Claude Code skills across directly, then had Claude rewrite the ones built for local PC files, since Gemini skills can't run scripts that need internet access. Users can create unlimited Skills but only 100 can be active at once; Skills don't yet work with Deep Research, Guided Learning, Canvas, or the video and music creators, pre-made Google Labs Gems won't migrate, and Google says sharing and Drive-file support are coming in the following weeks. Workspace accounts move later — business and nonprofit accounts keep Gems in Settings until March 2027, education until June 2027.

Who it matters to: Anyone with saved Gems on a personal account. Spark itself didn't change, but in 15 days the custom setups feeding it stop being editable, and a month later the originals are gone — oversized or oddly formatted Gems are the ones auto-conversion can quietly damage.

What you can do: This week, not on November 16: open each Gem you depend on, copy its instructions into a plain SKILL.md file you keep yourself, and download any attachments over the 100 MB cap or in unsupported formats. That one file then works in Gemini's new Skills — and, because it's the shared format, in Claude as well.

Learn more: XDA's hands-on Gemini Skills migration test, summarized with the key dates

Also worth knowing

• Unconfirmed — two flagship leaks in one night, neither company has confirmed anything. Screenshots posted November 1 show a "gpt-6.2-astra" label in Codex usage data, and a screenshot of Anthropic's Mythos page source code shows "Claude Mythos 5.5." An evidence review by Van Data Team rates the Astra screenshots plausible but untraceable to a full request, and the companion "Fable 5.5" claim weak — its price-list entries appear copied from Fable 5.1 and were removed a day later. Treat both as leaks to watch, not launches; for context, Gemini 4 Carbon currently tops the Artificial Analysis Intelligence Index v5 at 59, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57.
• OpenAI is reportedly planning a post-midterms push — including renaming data centers "AI compute campuses." TechCrunch, citing Axios, reports chief global affairs officer Chris Lehane previewing a campaign around AI "speed limits," U.S.-led global standards, and reframing data centers as "AI compute campuses," calling community debate over the projects "good and healthy." It lands days after a federal judge let the NAACP's pollution suit over xAI's Colossus turbines proceed — the naming fight and the court fight are the same fight.
• Anthropic clarified what Claude will and won't do on two fronts, per The Neuron's November 2 roundup: a new policy against sustained, needless abuse of Claude targets extreme, purposeless mistreatment — not frustration, criticism, creative writing or research — with ending the conversation as the primary response; and Anthropic confirmed it adjusted Claude Code's Opus 5.5 instructions to steer the model away from reproducing someone else's specific video game, while original and genre-inspired games remain allowed.
• Watermarking is becoming the default across labs. Alongside Anthropic's November 3 expansion, xAI will reportedly begin watermarking text generated through the Grok API by default on November 7 — two of the major labs marking output within the same week, both driven by the same transparency rules.

One practical takeaway

Two deadlines, one habit: export your setups before the platforms move them. Gemini's Gems freeze on November 17 and ChatGPT and Codex merge into one token budget in December — in both cases, the people who keep a copy of their own instructions and a baseline of their own usage will notice immediately if something breaks or shrinks, and everyone else will find out mid-task. Spend twenty minutes this week saving each Gem as a SKILL.md file (it doubles as a Claude skill) and screenshotting your Settings → Usage split. Future-you, in December, will know exactly what changed.

Sources

CNBC (agent blocks by JPMorgan Chase, Charles Schwab, Amazon, Kohl's and Neiman Marcus; Billions Network audit of 51 retailers; Accenture's Michael Abbott) via Daily Agentic's summary; The Wall Street Journal (Muse shopping pressure on Amazon) via Seeking Alpha/TradingView; The Information (Meta's Max humanoid with Fourier; Meta spokesperson statement) via Notebookcheck and Reuters; Apple China Newsroom and Alibaba via Xinhua and China Daily (Apple Intelligence in China, Qwen integration, November 4 beta and December iOS 27.3 release) via 9to5Mac, MacRumors and AF.net; OpenAI Help Center, "Understanding ChatGPT and Codex usage limits," updated October 31 (shared ChatGPT-Codex allowance) via 7MinAI, AI/TLDR and The Neuron; Anthropic customer email (November 3 text-watermark expansion) via The Decoder and Mashable, summarized by Madrobot and Superpower Daily; XDA Developers' November 1 hands-on (Gemini Gems-to-Skills dates, SKILL.md portability, 100 MB cap) via BlazeTrends; Van Data Team's evidence review (GPT-6.2 Astra and Mythos 5.5 leaks — unconfirmed); Axios (OpenAI "AI compute campuses") via TechCrunch. Summaries are original; follow the links for full reporting.