AI in 15 — November 07, 2026

An independent evaluator spent eleven weeks inside Anthropic and concluded that AI is now roughly doubling the speed of Anthropic's own AI research. The catch is that Anthropic redacted most of the evidence the evaluator found most convincing.

Welcome to AI in 15 for Saturday, November 7, 2026. I'm Kate, your host.

And I'm Marcus, your co-host. Happy Saturday, Kate. Today's top story is a report where the most important pages are blacked out, so my favorite question will get a workout.

"Can anyone check this?" I'll keep count. Here's what's coming up. METR measures whether AI is speeding up AI research at Anthropic.

Anthropic's IPO price talk drops after a record one-day tech rout.

Google makes AI video longer, cheaper and editable.

A tiny "decision model" nearly matches Claude Opus for a fraction of the price.

Plus a sandbox escape, a ChatGPT connector mix-up, Claude reading your bank account, and a GitHub repo that only accepts issues from your agent.

Marcus, on Wednesday we covered Anthropic's IPO filing, which warned that future models may "imminently" cross its threshold for AI-driven acceleration. Now someone outside has actually measured it. Who's METR, and what did they find?

METR is an independent group that evaluates AI models. About two and a half full-time staff worked inside Anthropic from August twelfth to October twenty-ninth, looking at one question: is AI speeding up AI research itself? Their yardstick was Anthropic's internal version of the Epoch Capabilities Index, which is basically a combined score of how capable the models are. Their conclusion is that Anthropic's capabilities are now improving about one point nine times faster than before.

Nearly double.

Right. The trend went from about fifteen points a year to about twenty-nine, so two years of past progress in one. The confidence interval runs from one and a half to two point eight times, so even the low end is a big jump. METR puts about eighty percent of that speedup down to AI tools, and says it found no convincing bottleneck that would slow things down over the next twelve months.

And the staff themselves? I imagine they have opinions.

Anthropic staff self-reported about a twelve-times productivity boost from the most capable internal model. METR calls that "very loose evidence," which is a polite way of saying people are bad at estimating their own productivity. I'd put more weight on the one point nine.

Okay, so here's the redaction part.

It's a big caveat. Anthropic could redact anything it wanted, and METR says it removed "most" of the evidence METR found most informative. That includes forecasts for future models and what METR calls the "signs of life" evidence for autonomous AI research. Anthropic says those sections contain confidential information that could help other AI developers. METR also didn't get query access to unreleased models. A planned check on how compute is split between capabilities work and safety work was cancelled partway through, at Anthropic's request.

Cancelled at the company's request. That's not a great look.

It isn't. And one more twist: METR says it leaned heavily on Claude itself to gather and analyze the data. So we have an AI helping measure how much AI is helping build that AI.

That's a lot of mirrors. So what should people take away?

Two things, and both are true. This is the first independent test with inside access of the "AI improving AI" loop at a frontier lab, and that's real progress on transparency. But the most important parts are things we have to take on trust. Ajeya Cotra, a well-known AI forecaster, didn't see the unredacted report either. Her read is that Anthropic has likely crossed the "dramatic acceleration" threshold in its own Responsible Scaling Policy. And if the speedup keeps compounding, capability doubling times could shrink from six months to three, and then to six weeks.

Six weeks. What did researchers on Hacker News think?

Mostly that experiments really do run faster now, which matches what people see day to day. The pushback was that a benchmark-heavy index might not reflect real-world progress. That's fair. Faster improvement on a score is not the same as faster improvement at useful work.

And it lands days before the roadshow.

Which is why I'd like the redacted parts shown to at least one more independent group. "Trust us, it's working" is a strange thing to price a company on.

Quick hits, and we'll stay with Anthropic. Yesterday we talked about the market slide squeezing its IPO. Now there's price talk. Marcus, how bad is it?

According to Bloomberg, prospective buyers are now discussing one point two to one point three trillion dollars, and some are floating as low as one point one. That's down from the one point four to one point five trillion investors had signaled to bankers, and well below the roughly two trillion talk in early October. HSBC's Max Kettner said "around one trillion could be a more comfortable level."

So up to seven or eight hundred billion dollars of imagined value, gone in a month.

On paper, yes. And a quick correction on the numbers. Yesterday we said the Nasdaq fell about five and a half percent on Thursday. That's the Nasdaq Composite. The Nasdaq 100, the big-tech index, fell six point four percent, and Bloomberg says that erased a record two point three trillion dollars in one day. A US chip index fell twelve percent, its worst day since March 2020. Nvidia lost almost four hundred and fifty billion dollars in market value.

All because of the midterms?

Mostly. Investors fear a slower, more regulated data-center buildout, and voters approved local measures that restrict data centers. Stocks bounced on Friday, with the Nasdaq 100 up about two percent by late morning. And to be fair to Anthropic, its revenue run rate reportedly topped eighty billion dollars at the end of October. One investor said a valuation near one point two trillion would have no trouble finding buyers.

Does the timeline hold?

The roadshow is due next week, with trading targeted before Thanksgiving. Reuters sources say investors are asking whether that still works. There's also a usage angle on Hacker News. Several developers said cheap local models like Qwen 3.8 are eating into what they pay for Claude, while others said local models still fall well short of Claude Code. That debate is really what investors are pricing: how much of the revenue is defensible.

Speaking of raising money during a rout. Thinking Machines, whose Inkling 2 Small model we covered yesterday, is reportedly raising five billion dollars.

Business Insider says the round is led by Accel at a forty billion dollar pre-money valuation. That's more than triple last year's roughly twelve billion. Nvidia is reportedly in talks to put in around two and a half billion. The weights are still promised for next week under Apache 2.0, after outside red-teaming.

Raising five billion the same week chip stocks had their worst day in six years takes some nerve.

Or a very good pitch. An American lab promising strong open weights in a field led by Qwen, DeepSeek and, this week, Baidu's new two-hundred-and-eighty-billion-parameter ERNIE 5.5 Flash Base is an easy story to sell. It still has to ship.

Google has a new video model, Gemini Omni 1.2 Flash. What's new?

Clips up to fifteen seconds, up from ten, adjustable one second at a time. You can feed it reference images for characters, products and style, plus short motion clips and audio cues, and it tries to keep faces, clothes and brand colors consistent across shots. The headline feature is multi-turn editing. You can say "change the lighting" or "make it vertical" without starting again from scratch.

That's the bit I'd actually use. What does it cost?

Seven cents a second at 720p, thirty percent cheaper than the previous version. Drafts render at 360p or 720p and upscale to 1080p or 4K. It's generally available in the Gemini API, AI Studio and Google Cloud, and it's built into Flow, Google Vids and Google Ads. Everything carries SynthID watermarks. Google lists no formal benchmarks, and it admits the model struggles with fast sports action and complex physics.

Built into Google Ads tells you who it's for.

Exactly. This is aimed at marketers making hundreds of variations, not creators making demos. Meanwhile, Hacker News mostly argued about naming. Gemini models are on version four, but this one is "Omni one point two."

Naming things remains the hardest problem in computer science.

This next one went viral on Hacker News, over four hundred points. Hightouch, a customer data company, tested what it calls "decision models." Marcus, explain.

They're small models that don't write text. You give them a condition and they return a confidence score between zero and one that it's true. Hightouch tested two, Jev and OpenAI's GPT-6 Luna, on identity resolution, which means spotting customer profiles that wrongly merged two different people. On a thousand-identity test, Claude Opus 5.5 scored ninety-nine percent accuracy for three dollars and eighty-six cents. Jev alone scored ninety-five point seven percent for five cents.

Five cents versus nearly four dollars.

And the clever part is the hybrid. Jev handles the cases it's confident about and sends the uncertain ones, scoring between thirty and ninety percent, up to Opus. That got ninety-eight point eight percent for eighty-eight cents. On a harder test set the hybrid actually beat Opus alone, ninety-five point three to ninety-five point one, at about fifty-five percent of the cost.

How does a cheaper setup beat the expensive one?

Calibration. When Jev is wrong, it usually knows it's unsure, so the hard cases end up with the right model. Think of a triage nurse who knows exactly when to call the specialist. Caveats from the thread: this is one narrow task, not general coding, and one commenter disputed the calibration analysis. But lots of people described similar setups, where a small local model does most of the work and only asks a frontier model when it's stuck.

Which loops back to that IPO question.

It does. Every request a small model handles is one the big labs don't bill for.

Security time. Two stories about the plumbing under AI agents. First, a sandbox escape.

Firecracker is Amazon's lightweight virtual machine technology, and a lot of AI agents run their code inside it. PWN Research published part two of a series showing how to turn a guest's access to host memory into root code execution on the host. The trick is to hijack a virtual CPU thread as it returns from the kernel's virtualization layer.

Should people running agents be worried?

Less than the headline suggests. The test ran without the "jailer," Firecracker's production containment that drops privileges and filters system calls, and commenters note the underlying flaw was disclosed weeks ago. So this shows how to exploit a known bug, not a new hole. Still, it's a reminder to run the jailer.

And the second one is about ChatGPT.

A developer had two connectors pointing at the same MCP server. MCP, or Model Context Protocol, is the standard way AI assistants plug into outside tools. ChatGPT silently sent a follow-up tool call through a different connector than the one the user had approved. That crosses a consent boundary, because each connector can carry different permissions. The server rejected the call, so no harm was done this time.

Did they report it to OpenAI's bug bounty?

They declined, because the Bugcrowd program required an ID card and biometric verification. Simon Willison pointed out that rule exists to stop floods of low-quality submissions. That's understandable, but it's also exactly how you lose reports from careful people. As agents run code and chain tools, the sandboxes and permission layers are the attack surface.

And while we're on OpenAI security, one follow-up on GPT-6 Cyber from yesterday.

The access rules are strict. Organizations need single sign-on, multi-factor authentication and IP allowlisting. Individuals need identity and liveness checks, a legal attestation and a physical FIDO2 security key. It also refuses to help with industrial controls, essential services and election systems. A model that has to be kept behind a hardware key says something about what it can do.

Product round-up. Marcus, Claude wants to look at my bank account.

Claude Personal Finance, in beta for US Pro, Max and Ultra subscribers. You link accounts through Plaid and ask about spending, bills or odd transactions. It's read-only, covers up to two years of history, and these chats are excluded from training. It had leaked as "Claude Money" back in September, and ChatGPT shipped a similar Plaid preview in May. We got the launch details from a newsletter, not the post itself.

Read-only is the right default. I'm not ready for it to pay my bills.

Neither am I. Perplexity launched Teammates, always-on work agents that juggle several assignments and that you can inspect or redirect, on the web for Max subscribers. Cursor's cloud agent can now build, host and secure live web apps from plain-language instructions, and it asks for approval before deploying. Mistral put managed agents, a visual workflow builder and an Agents SDK into public preview, with private-cloud and on-premises deployment due late November. And Microsoft's Foundry Local now runs Alibaba's Qwen 3.8 twenty-seven B on your device as a local coding model for GitHub Copilot.

So agents in my finances, my work queue and my deployments.

Running in the background, yes. Which is why the security stories matter.

Lightning round.

OpenAI opened outbound phone calling through its voice API to all organizations, and Hacker News immediately started worrying about robocalls. The Rust UI framework Rune doesn't accept pull requests. It asks you to "have your agent open an issue," and its own agents write the fix. Microsoft is winding down its AI cloud rental deal with Oracle, though commenters say that covered older A100 chips running Bing. Google is building a seventeen billion dollar data center in Wyoming, powered by an on-site fuel-cell plant. And Philippe Jaenada, who just won France's Prix Goncourt, admitted he used ChatGPT three times while writing the book, including for one sentence.

One sentence. I'd love to know which.

He's not saying. Smart man.

One to watch: where Anthropic sets its price range next week. If it lands near one point two trillion, the AI trade has steadied. If it slips toward one trillion, Thursday was more than a one-day scare.

Agreed, though I'd watch Nvidia's earnings on November seventeenth just as closely. That's the real test of whether the buildout is slowing.

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