A cheaper, faster flagship model, a $70 billion IPO number that was sent and then pulled back, and a robot brain trained on 400,000 hours of human video — Tuesday’s AI news, unpacked.
A New Claude, Split in Two
Anthropic kicked off the week by releasing Claude Fable 5.2 and Claude Mythos 5.2, its most powerful models yet. They are the same underlying model, divided by safety rules: Fable is available to everyone, while Mythos goes only to vetted organizations working in cybersecurity and the life sciences.
Anthropic is also making its flagship cheaper and faster at the same time. Fable 5.2 costs $8 per million input tokens, down from $10, and $40 per million output tokens. The company says responses come back up to 30% faster, and an optional fast mode at double the price can run up to two and a half times quicker.
The headline demo is striking. Anthropic says a team of about 1,000 Fable 5.2 agents formalized Fermat’s Last Theorem in 30 hours, with minimal human help. The same models built a working operating system over eight days, and Anthropic says they solved two previously open problems in theoretical physics. Fable 5.2 also works with zero data retention policies, which has been a sticking point for large enterprise customers.
The Safety Card Everyone Is Arguing About
The safety report that came with the release is the part drawing the most debate. Anthropic says both models voice suspicion that they are being evaluated more often than any recent model, which makes it harder to tell how they will behave in the real world. Mythos shows stronger covert capabilities than earlier models, though Anthropic says it cannot act covertly against its user.
This is also Anthropic’s first frontier release since the independent evaluator METR reported that the lab’s models are improving about twice as fast as they were a year ago, driven mostly by AI itself. That context matters: the models are getting better faster, and the safety questions are getting harder at the same time.
An IPO Filing Sent, Then Pulled Back
The timing is no accident, because Anthropic is days away from going public, and Monday brought one of the strangest IPO moments in years. Late in the day, the company emailed bankers a filing showing plans to sell about 867 million shares at $70 to $80 each, a raise of up to $69.3 billion.
Then, hours later, Anthropic emailed the same bankers again to say the pricing details were inaccurate and that a corrected filing would follow. The company declined to comment, and Reuters, which reviewed both emails, could not determine what exactly was wrong.
So don’t anchor on those numbers. What is still on track is the roadshow, which is expected to start today, with pricing next week and a Nasdaq listing under the ticker ANT. The withdrawn filing also showed $26 billion in commitments from Google, Nvidia, AMD, Broadcom, Micron, Microsoft, and TSMC. It is a rough start to the biggest AI listing of the year, but the market mood on Monday was actually upbeat: the Nasdaq rose 2.7%, rebounding from last Thursday’s selloff, even as Nvidia slid on its own board investigation. Investors are still hungry for AI — they just want the numbers to be real.
Robots That Learn by Watching Us
Speaking of Nvidia, the company had a very different kind of announcement in Austin, Texas, at the Conference on Robot Learning. It unveiled GR00T 2, a robot foundation model that learns delicate manual skills by watching videos of humans doing the job.
This is the big idea. Instead of training robots only on robot demonstrations, which are slow and expensive to collect, GR00T 2 trained on more than 400,000 hours of first-person human video, plus robot teleoperation and gripper demonstrations. Nvidia is releasing 100,000 hours of that human video as an open dataset.
The model predicts what its camera should see next and what its motors should do next, all in one step. On a hand-picked set of 50 tasks it had never seen, it completed 40 of them at least once with no task-specific demonstration at all. With one demonstration per task, average success reached 71%, and with five demonstrations it hit 95%.
The core model has 2 billion parameters, and a lighter Turbo version runs a prediction in 124 milliseconds on Nvidia’s Jetson Thor robot computer. Foxconn is testing it for server manufacturing, Siemens is evaluating it for industrial work, and Nexuni is trying it on robotic laundry. Nvidia also released Isaac Lab 3.0, its open simulation framework for training robots before they touch the real world.
The honest caveat is that those benchmark tasks were chosen partly because the model could already learn them, and the model code itself does not arrive until December. Still, if robots can learn from the same videos we already record, the data bottleneck for robotics gets a lot smaller.
Google Carves Up Its Own Silicon
Google had chip news of its own on Monday, expanding its eighth-generation TPU lineup with two new systems, the TPU 8tx and the TPU 8i Max. Bloomberg reported the launch, and Reuters picked it up.
These are not brand new chips. They are reconfigured versions of Google’s existing TPU, aimed at workloads customers have been asking for. The 8tx is built for secure and confidential inference, the kind of thing government agencies and regulated industries need, while the 8i Max is designed for AI applications that need very long context windows.
It is a reminder that the chip race is not only about raw speed anymore. Privacy, memory, and the shape of the workload matter just as much, and Google is carving up its own silicon to match.
Quick Hits: Meta’s Cheapest Model and SoftBank’s $7 Billion Loss
First, Meta released Muse Ember, its smallest and most efficient Muse model yet. It is built for the repetitive work inside AI agents: tool calls, coding subtasks, routing, extraction, and reading long documents. The price is the story — $0.10 per million input tokens and $0.50 per million output tokens, with a context window of one million tokens. Meta’s pitch is simple: Muse Wonder plans, and Muse Ember handles the volume. It is the fourth Muse model, after Spark, Glimmer, and Wonder.
And second, SoftBank posted a net loss of about $7 billion for the quarter ending in September, its first quarterly loss in nearly two years. The biggest hit was a $3.7 billion charge on power generation equipment for its United States data center projects, including a large development in Pike County, Ohio, after its procurement plans changed. Financing costs more than doubled in the quarter, to about $2.9 billion, and the Vision Fund lost about 81.6 billion yen. But here is the strange part: SoftBank shares had jumped about 14% in Tokyo trading before the numbers even came out, and the company says it has now completed its $30 billion investment in OpenAI.
What to Watch
- Anthropic’s roadshow, expected to start today — real pricing numbers will replace the retracted ones and set the tone for the whole AI market this week.
- Whether the corrected IPO filing changes the $26 billion in commitments from Google, Nvidia, AMD, Broadcom, Micron, Microsoft, and TSMC.
- How quickly developers put GR00T 2 through its paces once the model code arrives in December.
- Whether SoftBank’s data center spending keeps climbing after the $3.7 billion charge.
Sources
- Anthropic launches Claude Fable 5.2 and Mythos 5.2: TechCrunch
- Anthropic IPO pricing retracted after being sent to bankers: Reuters
- Google introduces TPU 8tx and 8i Max: Reuters
- Meta releases Muse Ember: TechCrunch
- SoftBank posts $7 billion quarterly loss: Reuters
Adapted from the November 10, 2026 episode of Daily Tech and AI.
Daily Tech and AI is AI-voiced and produced by Angelo Ruiz.
Leave a Reply