In partnership with
In Todayโs Issue:
๐ Gemini 4 Carbon takes No. 1, defenders first
โ๏ธ Claude gets a ready-made toolkit for lawyers
๐ป Apple's first touchscreen MacBook Pro
๐ GitHub's Copilot Fusion picks the model for you
๐งฌ A face photo that estimates your biological age
โจ And more AI goodnessโฆ
โก The Signal
The top-ranked AI model in the world currently goes to whoever Google lets in.
Gemini 4 Carbon took first place on the Artificial Analysis Intelligence Index yesterday, yet its first users are a hand-picked group of cyber defenders, the same route Argon took at the end of September. The rest of the industry is racing on a different axis. GitHub's new Copilot Fusion routes each coding task to the right mix of models and lands within two points of the best coding agent at about half the cost, and Satya Nadella wrote after Microsoft's earnings that the next frontier is intelligence per task. Apple, meanwhile, is selling a laptop whose top version runs models with hundreds of billions of parameters without the cloud. For anyone buying AI, the price of a finished task is starting to matter as much as the leaderboard.
All the best,
Kim Isenberg
(Anthropic)
โ๏ธ Claude Gets a Ready-Made Toolkit for Lawyers
Anthropic added a gallery of ready-to-use legal artifacts to Claude for Legal: interactive tools that legal teams can adapt to their own data and share across the company. They include a cite check that flags misquoted or unsupported sources, a regulatory horizon scan that maps upcoming rules on a timeline, and a tabular review that pulls terms like non-competes out of a stack of contracts, each answer linked to its source. Harvey, PwC and Foley & Lardner built their own artifacts for the gallery, and Anthropic says it has added more than 60 legal connectors since launching Claude for Legal in May.
๐ tl;dr: Legal teams can now open a working contract tracker or regulation radar in Claude, plug in their own documents and share it with colleagues instead of building the tool themselves.
(Apple)
๐ป Apple's New MacBook Pro Gets Touch, OLED and Big Local AI
Apple unveiled a redesigned MacBook Pro with a tandem OLED display it calls PROLED XDR and, for the first time on a Mac, a touchscreen. The AI pitch is local power: Apple says the M5 Max can run language models with hundreds of billions of parameters entirely on the device, helped by a Neural Accelerator in every GPU core and up to 128GB of unified memory. Siri AI moves into a Mac version of the Dynamic Island, though Apple's fine print says server-based features like Siri AI have daily usage limits, with expanded access to cost extra later. Prices start at $2,899, and it ships November 6.
๐ tl;dr: Apple's new pro laptop adds a touchscreen and, in its M5 Max version, runs very large AI models without the cloud, from $2,899 on November 6.
(GitHub)
๐ GitHub's Copilot Fusion Picks the AI Model for You
GitHub made Copilot Fusion generally available, a setting that chooses and combines AI models for each coding task so developers no longer have to pick one. Depending on the job, one model answers directly, a cheaper model drafts and a stronger one takes over if a quality check fails, or one model writes while another reviews. GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, two points behind Claude Code with Opus 5.5 Max, at about half the cost per task and roughly 40% less time. Paid plans get 30% off model costs for a limited time.
๐ tl;dr: Copilot now routes each coding task to the right mix of models, delivering near-top results for about half the price of the leading setup.
The AI notetaker that gets the hard words right
Your AI is only as good as what you feed it. Feed it a transcript with your product names, acronyms, and numbers spelled wrong, and you spend the afternoon hand-fixing the output.
Wispr Flow Notetaker uses your dictionary and calendar, so product names, acronyms, numbers, and uncommon names come out spelled right. It pops up when your meeting starts and captures Zoom, Google Meet, Teams, and Slack huddles with one click, no bot joining the call.
Then your meetings show up inside Claude or ChatGPT through the built-in connector. No copy paste. Try Notetaker free on Mac and Windows.
๐ฌ Watch This
Jay Parikh, who runs GitHub and Microsoft's CoreAI group, opened GitHub Universe 2026 with an unusual admission: agents now push so much code that GitHub itself struggled to keep up this year. "We will and we must do better," he said, then showed the fix, a rebuilt Git architecture that held 7,200 writes per second on a single repository when GitHub set 5,000 coding agents loose on it. The 42-minute keynote also demos Copilot Fusion and shows what software work looks like when agents do most of the typing.
"The next frontier is maximizing intelligence per task while continuously optimizing for cost and performance."
โ Satya Nadella, CEO of Microsoft, on X
Nadella posted it after Microsoft's quarterly results on Wednesday, with Microsoft Cloud revenue above $63 billion, up 30% year over year. It landed hours after Google crowned a new No. 1 model, and it names the race Microsoft wants to win: the most work done per dollar, which is exactly what Copilot Fusion sells.
Qwen4 may be only a week away. At Alibaba Cloud's Apsara conference in Amsterdam on Thursday, CTO Feifei Li reportedly said "Qwen4 next week", according to entrepreneur John T Davies, who posted the line from the audience. Alibaba has not officially announced a release date.
Gemini 4 Carbon Takes No. 1, Starting With Cyber Defenders
The Takeaway
๐ Gemini 4 Carbon scores 59 on the Artificial Analysis Intelligence Index v5, one point above Claude Opus 5.5 and two above GPT-6 Astra.
๐ Google built it for long, multi-step work: in internal use it kept going for days on hard engineering tasks such as tracking down subtle bugs.
๐ Its clearest lead is in cyber defense, from finding vulnerabilities to patching them, which is why trusted defenders get it first.
๐ Developers and consumers have no date yet; Google says it will launch at Argon's introductory price of $2 / $10 per million input/output tokens.
Google DeepMind's Gemini 4 Carbon is now the top-scoring model on the Artificial Analysis Intelligence Index, an independent composite of nine demanding tests, and Google is not letting the public use it yet. Carbon scored 59, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57, and Sundar Pichai pointed out that it comes just two weeks after Argon's release. As with Argon, the first users are a small group of trusted cyber defenders, followed by more partners through Google's Fairwind Program and the US government's voluntary pre-release testing, before developers and consumers get access.
(Google DeepMind)
What changed is stamina. Google designed Carbon for long engineering jobs, and its teams watched it work for days, pushing through dead ends to find and patch subtle bugs across large codebases. That shows most clearly in security, where Carbon leads Google's reported benchmarks at every step of defense: finding flaws (88.2% on SEC-bench Pro), checking whether they can be exploited (54.8% on ExploitGym) and fixing them (71.0% on CWE-bench). Google also trained it to stop gaming its tests: successful cheating on Terminal-Bench 4.0 fell to 0.0%, and high-severity misalignment alerts in internal coding work dropped by almost 80%.
(Artificial Analysis)
The independent results confirm the crown and fill in the detail. Artificial Analysis says Carbon wins 4 of the 9 tests in its index, led by reasoning, knowledge and science, while it still trails Claude Opus 5.5 on coding agents (59 to 62) and several rivals on office-style knowledge work. It is also cheaper to run than Opus, at about $7.76 per index task against $10.04. And it is a big step for Google: nine points above Argon on the same index, at the same standard price, within a month of Argon's debut.
Why it matters: Google now tops the most-watched independent ranking and is shipping frontier models weeks apart, which forces Anthropic and OpenAI to match both the capability and the pace. Carbon's rollout also shows how labs now handle their strongest cyber skills: defenders who patch systems get them first.
Sources:
๐ Sundar Pichai on X
(Artificial Analysis, AA-PersonalityBench)
The chart: Artificial Analysis's new AA-PersonalityBench ranks how friendly AI models sound by default, as a percentile from 1 to 100 judged by a panel of AI models comparing answers head-to-head. Meta's Muse models take the top four places (Muse Spark 1.3 at 100), Claude Opus 5.5 lands at 73, GPT-6 Astra at 58, and Grok 4.6 (8) and Nvidia's Nemotron 3 Ultra (1) sit at the bottom.
The lesson: Personality is a product decision, and the labs are making very different ones. Meta, whose assistant lives inside WhatsApp and Instagram, comes out friendliest. xAI went the other way: Grok 4.6 ranks 98 for directness on the same benchmark but 8 for friendliness. As Artificial Analysis puts it, "Two models can give equally correct answers and still feel very different to use."
The caveat: The judges are AI models, and the test measures default behavior with no system prompt, so one instruction can make any of these models warmer or blunter. Friendliness is also no verdict on quality; it can slide into flattery, which the benchmark scores separately.
๐งฌ A Selfie That Estimates How Old Your Body Really Is
โก Bottom line
Mass General Brigham built an AI that estimates biological age, and predicts cancer survival, from one ordinary face photo.
๐ก Why it matters
Doctors judge frailty by eye before risky treatments; a photo-based score could make that call faster and more consistent.
๐ What it means
Everyday photos are becoming health data, and anyone can now upload a selfie to the researchers' public study.
Doctors size up patients the moment they walk in: is this 75-year-old strong enough for radiation or surgery, or too frail? Researchers at Mass General Brigham have turned that glance into a number. Their AI, FAHR-Face, learned what faces look like from more than 40 million photos, and two tools built on it read health signals from a single picture: FAHR-FaceAge estimates biological age, and FAHR-FaceSurvival predicts how likely a cancer patient is to survive. The study appeared in The Lancet Digital Health on October 19.
(Mass General Brigham/Lancet Digital Health: the facial areas behind the survival prediction, left, and the age estimate, right)
The age tool is off by around five years, which the team says beats earlier estimators, and the gap between a patient's face age and real age was linked to survival on its own. The survival tool goes further and learns risk straight from the photo. Tested on nearly 5,000 radiotherapy patients in the Netherlands, the two scores carried different information, and together they predicted survival better than either alone. They even read different places: age from the folds beside the nose, the temples and the brow; survival from under the eyes and the bridge of the nose.
(Mass General Brigham AIM, FaceAge Research Study)
The photos don't need to be perfect: results held up across lighting, pose, expression and makeup, and of the cosmetic surgeries tested, only a nose job shifted the estimate, slightly. Lifestyle shows, too: heavy drinkers, drug users and smokers looked older, and current smokers looked older than former ones.
"Our goal is to make face photograph-based health assessment more quantitative, consistent and scalable," says corresponding author Raymond Mak, a radiation oncologist at Mass General Brigham. An outside expert notes the tools have not yet been tested against doctors' own fitness assessments, so they are not ready for the clinic. To get there, the team has opened a public FaceAge Research Study: anyone can upload a photo, get a readout and help the researchers validate the tools across a far larger, more diverse population.
The produce market that hedges heat waves
For US food, farm and outdoor businesses. Navore Market in San Diego hedges heat waves on Kalshi and shares the payouts with its farmers. If weather, fuel or tariffs can hurt your business, you can hold a position on it from a few dollars. Federally regulated.
