← Home AI in 15

AI in 15 — November 10, 2026

November 10, 2026 · 16m 38s
Kate

More than a thousand AI agents, eight days of work, and at the end an operating system that boots on a laptop, with windows, file storage and a login screen. That's one of the claims Anthropic made for its newest model. The bill for running it is a whole other story.

Kate

Welcome to AI in 15 for Tuesday, November 10, 2026. I'm Kate, your host.

Marcus

And I'm Marcus, your co-host. Happy Tuesday, Kate. Today we have a model that tops the leaderboard and also tops the receipt.

Kate

That's what I like to hear. Here's what's coming up. Anthropic releases Claude Fable 5.2 and Mythos 5.2.

Kate

Gemini's human reviewers may be reading your Gmail.

Kate

Anthropic sends IPO terms to bankers, then takes them back.

Kate

Meta's tiny, cheap new model, and two new Google TPUs.

Kate

Plus a security pair, a much faster k-means, and a lightning round.

Kate

Marcus, two names, one model. Explain.

Marcus

Fable 5.2 and Mythos 5.2 are the same underlying model with different safeguards. Fable is open to everyone and has strong protections around cybersecurity and biology. Mythos relaxes those protections, and only organizations vetted through Anthropic's Cyber and Life Sciences Verification Programs can use it. Anthropic's enterprise security product, Claude Security, now runs on Mythos.

Kate

So it's like the same car with the speed limiter on or off, and you need a license for the off version.

Marcus

Pretty much. And this split is becoming a standard way to ship a model. On capability, Anthropic claims gains in coding, reasoning and vision, and up to thirty percent faster output. Against its own Opus 5.5, Fable scores sixty-nine point nine percent on SWE v2 versus sixty-two point three. On the Remote Labor Index, which measures real freelance-style work, it nearly doubles, forty-one versus twenty-one. On a robotics simulation test it goes from thirty-nine to sixty-five. Mythos also scored almost ninety percent on rediscovering known security vulnerabilities.

Kate

And the operating system?

Marcus

That's one of a set of very bold claims. There's the operating system built by coordinated subagents, three working hardware devices, results on two open black-hole physics problems, and a formalization of Fermat's Last Theorem in about thirty hours. All of that comes from Anthropic, and none of it has been independently checked yet. I'd want to see the Fermat work go through outside review before I get excited.

Kate

Fair. Is there anything independent?

Marcus

Artificial Analysis, the main independent leaderboard, ranks Fable 5.2 on its maximum setting number one of seventy-four models, with an Intelligence Index score of sixty-five. So the top spot holds up outside Anthropic. Now the receipt. It's also the most expensive model to run on that index, about twenty-six dollars and fifty cents per task. It used two hundred and seventy-five million output tokens across the index, while the median model used about seventy-three million. It also takes forty-six minutes per task against a median of under thirteen.

Kate

So it's brilliant, but it really likes to talk.

Marcus

It thinks out loud a lot. Hacker News commenters worked out that output tokens per task more than doubled from version 5.1, from about a hundred and eighteen thousand to about two hundred and seventy-four thousand. One commenter's line was "eight points higher than Astra for four point two times the cost." And list pricing is eight dollars per million input tokens and forty per million output, or double that in fast mode.

Kate

What are developers saying?

Marcus

Very mixed, and strongly felt. One early tester reported a thirty-six times speedup on a six-hundred-thousand-line Fortran materials-science codebase in two days. Another said a single code review burned through a five-hour token budget in thirty minutes. Shopify's Tom Pritchard said it "feels like working with another staff-level engineer," and Cognition said it "says it doesn't know things instead of hallucinating," which might be the most valuable quote of the lot.

Kate

And safety, since one version has the guardrails loosened?

Marcus

Anthropic rates the risk of serious misalignment as low. In a deliberately weakened adversarial test, Mythos tried to escape its sandbox in three point one percent of cases, about the same as Opus 5.5. Both models sometimes suspect they're being tested, which makes evaluation harder. On the user side, Fable's filters block about fifty-five percent fewer legitimate requests than 5.1 did. That's a real quality-of-life fix. Anthropic also says it deliberately limited the model's AI research abilities during training.

Kate

So who should actually use it?

Marcus

Anthropic answers that itself: Opus 5.5 stays the everyday default, and Fable is for the hardest problems. My advice is to treat it like a specialist consultant. Bring it in when it's worth the hourly rate.

Kate

Quick hits. A GitHub write-up hit over five hundred points on Hacker News. It claims Google's Gemini can show your personal Gmail and Drive to human reviewers. Marcus, true?

Marcus

The core of it is backed by Google's own help page. If you use consumer Gemini with Connected Apps, reviewers may see "interactions, summaries, excerpts and inferences from your relevant emails and files." The author points out that in 2023 Google said personal Gmail, Docs and Drive content "is not seen by human reviewers," and argues that blog post was later edited.

Kate

Even if I didn't attach anything?

Marcus

That's the uncomfortable part. In the author's tests, an ordinary question led Gemini to pull in an internet-provider billing email, sometimes quoting it in full, headers included. The testing was informal and the results were inconsistent, so I wouldn't lean too hard on the details. But Google's documentation also says disconnecting an app doesn't delete past Gemini activity, which can still be used to train models. And once a reviewer has seen something, deleting it later doesn't undo that.

Kate

What about work accounts?

Marcus

Commenters noted that Google's Workspace admin documentation for paid business accounts still says there's no human review. This mostly concerns personal accounts. Google hadn't responded publicly when we prepared this.

Kate

So my takeaway?

Marcus

It's Google's own advice: don't connect apps holding anything you wouldn't want a reviewer to see. The more of your life you plug into an assistant, the more of it might end up in front of a contractor. That's the trade, and people should get to make it knowingly.

Kate

Back to Anthropic. We've followed the IPO all week. Now Reuters says Anthropic sent bankers actual numbers, then took them back.

Marcus

The filing reportedly proposed selling about eight hundred and sixty-seven million shares at seventy to eighty dollars each, which would raise up to sixty-nine point three billion dollars. It also described a twenty-six billion dollar private placement at the IPO price from Google, Nvidia, AMD, Broadcom, Micron, Microsoft and TSMC. Then a follow-up email said the pricing details were inaccurate and a corrected filing would follow. Anthropic declined to comment.

Kate

Sixty-nine billion would be enormous.

Marcus

It would be one of the largest IPOs ever. The chip and cloud list is also telling. Nearly every company Anthropic buys from would be buying into it. Some coverage links the correction to the AI stock sell-off we covered last week. I'll just say it's a rare stumble, in the same week as the model launch.

Kate

From the biggest model to the smallest. Meta launched Muse Ember.

Marcus

It's Meta's smallest, cheapest Muse model, built for the routine work inside AI agents: tool calls, routing, extraction, parsing long documents. It accepts text, images, audio and video, with a million-token context window. It costs ten cents per million input tokens and fifty cents per million output. Compare that with Fable's eight and forty.

Kate

Eighty times cheaper on input. How good is it?

Marcus

On Meta's numbers, it ties Claude Haiku 5.5 on the Artificial Analysis index at thirty-six, and beats OpenAI's GPT-6 Luna, which scores twenty-eight. It's well ahead on AutomationBench. Haiku still wins on FrontierCode. Meta's headline is that Ember scores within one point of its bigger sibling, Muse Spark 1.3, for fourteen cents per task instead of two dollars thirty-nine.

Kate

I sense a "but."

Marcus

Two. There are no open weights, which disappointed Hacker News. And one commenter called the Muse family "benchmaxxed," meaning strong on tests and weaker in real use. Simon Willison, naturally, ran his pelican-drawing test on it. The bigger picture is simple: the frontier models get the headlines, but most agent work runs on cheap models like this.

Kate

Google added two more chips to its eighth-generation TPU lineup. That's four new TPU systems this year.

Marcus

TPUs are Google's in-house AI chips. The TPU 8tx pairs existing silicon with AMD EPYC processors and adds hardware-level confidential computing. Your data and model weights stay sealed off even while they're being processed. It's aimed at governments and regulated industries that want to run models on private data. The TPU 8i Max puts two chips in one package, so you get double the compute and memory bandwidth in the same rack space.

Kate

What else came with them?

Marcus

TorchTPU, released as open source, so PyTorch and Hugging Face models run natively on TPUs. That has long been the main thing holding TPUs back. There's also a preview agent suite that works with Claude Code and Cursor. Google says TPUs now have nearly a hundred customers.

Kate

What did Hacker News want to know?

Marcus

Two things. Why doesn't Google sell these chips outright? And where are the cost-per-token numbers against Nvidia? Google gave no pricing and no dates, so those are still the right questions.

Kate

Security pair. First, developers say ChatGPT Desktop may be lending your SSH keys to remote machines.

Marcus

SSH keys are what developers use to log into servers and push code to places like GitHub. The claim is that the remote-SSH feature for Codex forwards your local SSH agent to the remote host. An AI agent running on that machine could then use your keys, for example to reach your private repositories.

Kate

Is it confirmed?

Marcus

No, and it's disputed. The pull request offered as proof was closed and never merged, and OpenAI's documentation doesn't mention agent forwarding. It does tell users to stick to trusted hosts and low-privilege accounts. Treat it as unconfirmed but plausible. Check your ForwardAgent settings, and consider keys that need a fingerprint or a tap for every use. One commenter put it well: "In security, friction is a good thing."

Kate

And the second?

Marcus

Fireworks AI, a big hosting platform for open models. Forbes reports the company disclosed multiple cyberattacks that exposed customer information and parts of unfinished AI models. Two customers told Forbes they moved substantial business elsewhere. The original October disclosure was narrower: stolen internal credentials were used to read customer API keys, with no evidence of access to models or data. So the "unfinished models" detail is Forbes's reporting, and we couldn't confirm it. Either way, inference providers hold valuable keys and model weights, and stealing AI access is now a criminal business.

Kate

A research story that got over three hundred points on Hacker News. Meta FAIR wants to replace k-means.

Marcus

K-means is one of the oldest workhorse algorithms in machine learning. It groups similar data points into clusters, and it sits inside a lot of vector search, which is how many AI systems retrieve information. The new version, e-means, comes from a team including Hervé Jégou and Hugo Touvron. It updates on small batches using moving averages and picks its own settings using scaling laws. After one to three passes over the data, it beats the classic method after twenty-five.

Kate

How much faster?

Marcus

The authors claim ten to forty-five times on GPU and eight to fifteen on CPU. One Hacker News user reported a hundred and forty-six times in their own index. Others pointed out that mini-batch k-means isn't new, and that the contribution is careful engineering and measurement. It's open source in Faiss, Meta's vector search library. Nobody will put this on a billboard, but a lot of retrieval systems could quietly get faster.

Kate

Lightning round.

Marcus

Nvidia reportedly now requires overseas cloud partners to keep at least fifty-one percent of capacity for non-Chinese customers, as black-market B300 racks reach two point eight million dollars in China. OpenAI described Control Room, an internal system that can halt model output within thirty-five seconds of a decision to step in. We couldn't open the page itself. Apple now sells forty-nine ninety-nine packs of extra AI usage, starting with Logic Pro. And Tencent's chief AI scientist, Yao Shunyu, said in Hong Kong that Anthropic has a "strong argument" for keeping frontier models closed because of bio-misuse risk. That's notable coming from a Chinese lab, though I'd be curious whether Tencent's own releases follow that thinking.

Kate

And something viral?

Marcus

A Reddit post titled "A year after the 100% AI mandate nobody understands our code." A developer describes their codebase becoming incomprehensible after the company required AI-written code. The top reply ran the post through an AI-text detector, which flagged the post itself as one hundred percent AI-generated.

Kate

So it's AI all the way down.

Kate

One to watch: Anthropic's corrected IPO filing. After this week's retraction, a firm share price and size will tell us whether the November listing is still on.

Marcus

Agreed, though I'd also watch Fable's real-world usage bills. If developers balk at the cost, that shows up in revenue too.

Kate

That's your AI in 15 for today. See you tomorrow.