AI News Today — November 6, 2026: Muse AI Gets a Face in the US, Claude Reads Your Bank Account Read-Only, and Anthropic's Model Progress Doubles

The big picture

In one day, the agents got a face, a wallet — read-only — and a warning label. Meta opened video calls with its Muse agent to everyone in the United States, Anthropic let Claude look at your real bank transactions without being able to touch the money, and Google made AI video longer and cheaper inside the same Gemini app that hosts Spark. Underneath the launches, the two most consequential stories were about evidence: an independent evaluator says Anthropic is now improving its models about twice as fast as before, mostly by using AI to build AI — and that the lab redacted most of the proof — while OpenAI kept its most capable cybersecurity model locked behind an approval process. Capability is still compounding. Access to it, and to the evidence behind it, is getting more carefully rationed.

What happened today

Muse AI: Muse Voice video chat opens to everyone in the US, and Muse launches in Mexico

What happened: Meta announced on November 5 that Muse Voice — the feature that lets you call and video-chat with its consumer AI agent, Muse — is now available to everyone in the United States. First shown at Meta's Connect conference in September, it runs on two new models, Muse Realtime Avatar and Muse Realtime Voice, which power an animated avatar nicknamed Jolly. You can customize the avatar to look like a person, an animal, or another character; as it speaks it uses expressions and gestures matched to the conversation's tone. Muse hears you but does not see you. Conversations work on mobile, web, and Mac now, and Meta says they will later come to its AI Glasses and the upcoming Muse Charm — the small wearable, expected before year-end, whose screen will carry the same avatar. The same week, Muse also launched in Mexico on November 5 — its fifth market and its first in Latin America, after the US (September 8), Canada (September 18), Japan (October 23), and Singapore (October 28) — for adults 18 and over, per availability tracking checked against Meta's official pages.

Who it matters to: Anyone tracking the personal-agent race. Muse is the first of the big personal agents to make a face-to-face video call its default way of talking, and it did so while its rivals are making the opposite interface bet — OpenAI is merging ChatGPT's Chat and Work tabs into one text workspace, and Google folded Spark into a Task toggle. Meta is betting presence beats productivity UI; the Mexico launch also shows it expanding the agent's footprint market by market rather than all at once.

What you can do: If you're in the US, you can try a Muse Voice call today from the Muse app on mobile, web, or Mac — use it for one real delegated task and judge whether talking to a face makes you more or less likely to check its work. If you're outside the five live markets, there is still no announced date for the UK, Europe, or India, and Meta treats VPN workarounds as a terms violation — wait for the official launch rather than risking the account.

Learn more: TechCrunch on the Muse Voice US rollout

Claude can now read your bank account — read-only, through Plaid

What happened: Anthropic launched Personal Finance in Claude on November 5, a beta rolling out this week to US subscribers on the Pro, Max, and Ultra plans. You link bank, credit-card, loan, and investment accounts through Plaid — you sign in to your bank on Plaid's screen, so Claude never sees your bank password — and then ask questions about your own money in plain language. For each account you connect, Claude can read the bank and account name, account type and last four digits, balances and credit limits, and up to two years of transactions, across up to 16 banks. The boundary is explicit and is the point of the launch: access is read-only — in Anthropic's words, Claude "can't move, manage, or invest your money" — the feature carries no ads or sponsored recommendations, Anthropic offers no credit products of its own alongside it, and it says the tool is not designed to give investment advice. Chats inside the Personal Finance hub are not used for model training. ChatGPT and Perplexity shipped similar Plaid-powered finance features earlier this year; Claude is the third major assistant to ask for your transactions.

Who it matters to: Anyone who has ever exported a bank CSV into a chatbot — which is to say, a large share of power users, who until now pasted their financial lives into tools with far weaker controls. A read-only, revocable, per-account connection through an established intermediary is a materially better pattern than uploads and screenshots. It also matters to banks and budgeting apps: the assistant is becoming the interface people query their money through.

What you can do: If you try it, connect one account first, ask it questions you already know the answer to — last month's spending by category, your recurring subscriptions — and check its numbers against your bank's own app before connecting more. Anthropic itself warns Claude can make mistakes in its answers and advises asking which transactions sit behind any figure. And remember the scope limit: the no-training promise covers chats inside the Personal Finance hub, so keep finance questions inside the hub.

Learn more: TechCrunch on Claude's Personal Finance tool

METR: Anthropic is improving its models about twice as fast — and redacted most of the evidence

What happened: METR, the nonprofit AI evaluator, published an independent assessment on November 5 after spending roughly 11 weeks (August 12 to October 29) inside Anthropic's internal data — benchmarks, training compute, research-efficiency measures, and staff interviews, with no payment taken. Its headline finding: since October 2025, progress on Anthropic's internal capability index has run at about 1.9 times its earlier rate (95% confidence interval 1.5 to 2.8 times) — roughly two years of historical progress compressed into one, about the distance from Claude 2.1 to Sonnet 4.5. METR is confident faster-growing training compute does not explain the jump and estimates that perhaps 80% of the current acceleration comes from AI itself helping with the research. The assessment also counted three models Anthropic has finished but not released, trained between July and October. The caveat is printed by METR itself, in unusually plain language: Anthropic "redacted a majority of the evidence we found most informative" from the roughly 100-page public report, which Anthropic describes as removing confidential information that could accelerate other developers. METR could publish its conclusions, but not most of its receipts.

Who it matters to: Everyone planning around model improvement — which now includes any business with an AI roadmap. If the pace of capability gains at a frontier lab has genuinely doubled because AI is doing much of the research labor, planning assumptions built on last year's rate of improvement are already stale. The redaction fight matters just as much: the strongest public claim about AI speeding up AI currently rests on evidence the public cannot inspect, and METR — the party with the most to gain from the report being believed — is the one flagging that.

What you can do: Treat the 1.9x figure as a well-sourced estimate, not a measured fact: it comes from a credible evaluator working from internal data, with a wide confidence interval and redacted support. Practically, shorten your own planning horizon — evaluate models and agent tools quarterly rather than annually, and avoid multi-year contracts priced on today's capability gap. If a vendor quotes acceleration numbers at you, ask the METR question: what evidence can I actually see?

Learn more: Analysis of METR's Anthropic acceleration assessment

OpenAI's GPT-6 Cyber found 948 Linux exploits — and you can't have it

What happened: OpenAI released GPT-6 Cyber on November 5, a model specialized for authorized security work: finding vulnerabilities, developing exploits, and proposing patches. OpenAI classifies it Critical for cybersecurity under its Preparedness Framework — its highest designation, up from High for GPT-5.6 Cyber — and is releasing it only through the API and Codex to a select group of approved Daybreak Red customers, with identity checks, organizational security requirements, and API IP allow-listing; OpenAI's documentation lists it at $20 per million input tokens and $100 per million output. The headline results are OpenAI's own: in an automated pipeline run against current Linux kernel code, the model worked from 2,308 candidate findings to reproduce 1,724 vulnerabilities and build working local-privilege-escalation exploits for 948 distinct flaws, and it found 81 previously unknown Firefox vulnerabilities that Mozilla classified as high severity. On OpenAI's ExploitGym benchmark it scored 48.5%, against 32.1% for GPT-5.6 Cyber — and, in the result OpenAI emphasizes most, it went out of bounds in under 1% of honeypot tests, down from about 5% for its predecessor. Those are laboratory figures from the model's maker, not independent measurements.

Who it matters to: Security teams and anyone responsible for patching. A gated model that can mass-produce working exploits for real kernel flaws compresses the time between a vulnerability existing and a vulnerability being weaponizable — by defenders first, if the gating holds, which is exactly the bet OpenAI is making. The less-than-1% out-of-bounds rate, if it holds up in the field, is also the first concrete progress metric on the agent-control problem this site has tracked all autumn.

What you can do: Unless you are an approved Daybreak Red customer, there is nothing to try — and that is the story. What you can do is defensive and immediate: the Linux and Firefox findings mean patches are coming or already landing, so make sure your kernel and browser update channels are actually current, and ask your security vendors how they are using — or defending against — exploit-capable models. If your organization does authorized testing, the Daybreak Red application is the only legitimate route in.

Learn more: TechCrunch on OpenAI's new cybersecurity model

ChatGPT Images 2.6 makes GIFs and generates 50% faster

What happened: OpenAI launched Images 2.6 on November 4 across ChatGPT, Codex, and the API, on all tiers. The headline addition is native GIF creation: turn a text prompt, a sketch, or an existing image into a short looping animation in one step, with frame-by-frame editing in the fuller rollout. OpenAI says generation is 50% faster than the previous version — a company claim, not an independent benchmark — and says its tools now produce more than 4 billion images a week. Two new API models split the work: GPT-Image-2.6 Flare, tuned for speed (OpenAI cites an image in as little as six seconds at low quality), and GPT-Image-2.6 Sunburst, tuned for detail and for keeping a person, product, or scene consistent across repeated edits. Early reports note the GIF feature is rolling out gradually and is missing on some accounts, sometimes producing frames without a finished animation.

Who it matters to: Creators, marketers, and small teams who produce social content. A shareable animation used to mean a video tool, a timeline, and an export; it is now a sentence typed into the same chat where the still image was made. The speed-versus-quality split in the API matters to developers the same way: pick Flare for drafts and volume, Sunburst for the version that ships.

What you can do: Try one real asset, not a test doodle: take a product photo or a mascot you already use, ask for a short GIF, and check the result the way your audience will — on a phone, at feed size, looking for warped text and drifting details. If your account doesn't have GIF creation yet, that's the staged rollout, not your settings; check again over the coming days rather than reinstalling anything.

Learn more: Digital Trends on ChatGPT Images 2.6 and GIF support

Gemini Spark: no new Spark announcement — Gemini video jumps to 15 seconds, and the Gems freeze is 11 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 beside the prompt box, multistep tasks that keep running in the cloud across connected apps, up to 15 tasks at once, for Google AI Pro and Ultra subscribers on personal accounts, everywhere Gemini Apps are supported except the EEA, the UK, Switzerland, and Nigeria. The Gemini news that did land on November 5 is video: Gemini Omni 1.2 Flash is now generally available through the Gemini API and Gemini for business, generating and editing clips of 3 to 15 seconds — up from a 10-second ceiling — with multi-turn conversational edits that hold the rest of the scene steady. Google cut the 720p price 30% to $0.07 per second, so a full 15-second clip costs about $1.05 at API rates; Google Vids users can make up to three 15-second videos a month free, and 15-second generation in the Gemini app itself is rolling out first to AI Ultra subscribers. Underneath it, the Spark-adjacent countdown tightened by a day: personal-account Gems become read-only on November 17 — 11 days from today — and are permanently deleted on December 17, with Google converting them to portable Skills files.

Who it matters to: Spark users first: nothing about your tasks changed today, but the setups around them are on a clock — anything you built as a Gem stops being editable in 11 days. Video creators and developers get the more immediate news: 15 seconds is the length of an actual social ad spot or product explainer, and a 30% price cut at that length changes the math on generating several variants and picking the best.

What you can do: Two separate checklists. For video: if you have API or Vids access, generate one 15-second version of something you previously had to split into two 10-second clips, and compare the consistency across the join — that is the specific thing this release claims to fix. For Gems: do the backup this week, while it is still boring — copy each Gem's instructions into a plain SKILL.md file you keep yourself, and download any attachments that won't convert. The file then works in Gemini's Skills and, because it is a shared format, in Claude as well.

Learn more: Coverage of Gemini Omni 1.2 Flash's 15-second video and price cut

Also worth knowing

• Palo Alto Networks is in talks to buy agent-security startup Clutch for about $500 million, TechCrunch reports — talks, not a closed deal. Read it next to the GPT-6 Cyber story above: securing agents and arming security agents are becoming the same market. TechCrunch's report
• Airbnb says its AI travel agent arrives in the first half of 2027. An intent announcement with no product to test yet — file it beside Uber's four agent-booked rides from yesterday's briefing as another marker of how early agent commerce actually is.
• Dropbox is acquiring Tigris Data for $350 million to bolster the storage infrastructure under its AI products — the unglamorous layer every AI feature quietly depends on.
• Box put Nvidia's Nemotron models inside agent-first content retrieval at BoxWorks 2026, alongside Box Architect, which proposes and builds workflows from a company's own content — both folded into its Enterprise Advanced plan.
• Switzerland's Embodied AI unveiled AIR, a wheeled humanoid with tendon-driven arms designed to learn factory tasks by watching workers, at Humanoid Europe 2026 in Munich on November 6; the company says tasks typically transition to autonomous operation within about a month, with CE marking planned for 2027 and testing underway at major manufacturers. The AI Insider's report
• OpenAI president Greg Brockman says the company could be IPO-ready in 2027, per TechCrunch — a readiness remark, not a filing, landing in the same week Anthropic's own S-1 has been the industry's favorite document.
• Anthropic paid for news for the first time, in a deal with the Associated Press that let Claude show live US election results with source attribution, Adweek reports via daily roundups — a small deal with a large precedent for how assistants license the news they summarize.
• Wizard Cyber launched CYBERSHIELD AI, an agentic security-operations platform where specialized agents carry triage and investigation while human analysts stay accountable for consequential decisions — live now for its managed-security customers on Microsoft Sentinel, in three packages.
• Aina's smart ring puts a screen, a joystick, and a microphone on your finger for talking to AI agents — the latest entry in the race, alongside Meta's coming Muse Charm, to find the wearable people will actually keep wearing.

One practical takeaway

Give AI read access before write access — and keep your own copies of everything. Today's launches drew the line in the same place three times: Claude can read two years of your transactions but cannot move a dollar; GPT-6 Cyber can build 948 working exploits but sits behind identity checks and IP allow-lists; METR could read Anthropic's internal data but could not publish most of what it found. Apply the pattern to your own stack this week: when a tool asks for a new connection, start with the read-only scope, one account, questions you can verify — and before November 17, make sure anything you've built inside someone else's AI product, starting with your Gemini Gems, also exists as a plain file on your own disk.

Sources

TechCrunch (Meta's Muse Voice US rollout — Muse Realtime Avatar and Muse Realtime Voice, mobile/web/Mac, AI Glasses and Muse Charm to follow, November 5; Claude Personal Finance launch, November 5; OpenAI's GPT-6 Cyber release and the questions around its safeguards, November 5; Palo Alto Networks' talks to acquire Clutch for about $500M, November 5; Airbnb's AI travel agent planned for the first half of 2027, November 5; Dropbox's $350M acquisition of Tigris Data, November 5; Box's Nemotron-powered agent retrieval at BoxWorks 2026, November 5; Greg Brockman's IPO-readiness remark, November 5; Meta's warning on data-center backlash, November 5); METR's independent assessment of AI R&D acceleration at Anthropic (1.9x rate, 95% CI 1.5–2.8x, ~80% AI-attributed estimate, three unreleased models, redactions of most informative evidence, November 5), via Madrobot's analysis and Pivot News' and Business Insider's reporting; Anthropic's Personal Finance announcement and Claude release notes, via Tiller, Unite.AI, and MLQ (Plaid connection, read-only scope, up to two years of transactions, up to 16 banks, Pro/Max/Ultra US beta, November 5); OpenAI's GPT-6 Cyber publications and system card, via MLQ, CybersecurityNews, and The Hindu BusinessLine (948 Linux privilege-escalation exploits from 2,308 candidates and 1,724 reproductions, 81 high-severity Firefox vulnerabilities, ExploitGym 48.5% vs 32.1%, under-1% out-of-bounds rate, Critical classification, Daybreak Red gating, November 4–5) and Apidog's check of OpenAI's Daybreak API documentation for the $20/$100 pricing (November 6); OpenAI's Images 2.6 announcement, via Digital Trends and EveryBeat (GIF creation, 50% faster generation as a company claim, more than 4 billion images a week, Flare and Sunburst API models, November 4); Google's Gemini Omni 1.2 Flash announcement, via Windows Report, Digit, and MobileTelco (general availability, 3–15 second clips, $0.07/second 720p — 30% below Omni 1.1 — three free 15-second videos a month in Google Vids, November 5); Muse availability tracking checked against Meta's official pages (Mexico launch November 5 as the fifth market; US September 8, Canada September 18, Japan October 23, Singapore October 28) and Unite.AI's Muse Voice technical summary (November 5); Adweek's report of Anthropic's first paid news deal with the AP, via eYssen's November 5 roundup, which also summarized Anthropic's 207,049-session Claude science-usage study (November 5); The AI Insider and BusinessWire (Embodied AI's AIR humanoid unveiling at Humanoid Europe 2026, November 6); Wizard Cyber's CYBERSHIELD AI launch release (November 6); Gemini Spark guides from Everpod and BetterClaw checked against Google's Gemini Apps Help pages (Task mode requirements, availability exclusions, and the Gems-to-Skills timeline — read-only November 17, deletion December 17). Summaries are original; follow the links for full reporting.