Anthropic launches Fable 5.2 and Mythos 5.2 just ahead of its IPO, while the Nvidia affair turns political and Australia plans an AI regulator.
Anthropic launches Claude Fable 5.2 and Mythos 5.2 with 20% lower API prices
Certain Models Products
Fable 5.2 is available on paid Claude plans, Claude Code, the developer platform and the major clouds; Mythos 5.2 is the same base model with different safeguards, limited to vetted cybersecurity and life-sciences users. Anthropic says output is up to 30% faster and API prices fall to $8 (input) and $40 (output) per million tokens. It also claims a team of 1,000 Fable agents formalized Fermat's Last Theorem in about 30 hours.
Anthropic IPO filing: up to $69.3B at $70-80 a share, debut possible week of November 16
Medium Business
According to a filing seen by Reuters, Anthropic plans to sell about 867 million shares on Nasdaq, with Google, Nvidia, AMD, Broadcom, Micron, Microsoft and TSMC together putting in $26 billion. Yahoo Finance notes that market turmoil could still affect the timing or the valuation (now $1-1.5 trillion rather than above $2 trillion).
Update: Nvidia affair brings political pressure on chip exports; Jeffries demands briefing, shares fall
High Policy Chips Business
House Democratic leader Jeffries wants a full briefing and backs writing AI chip export limits into the defense bill. Per Seoul Economic Daily, Nvidia fell as much as 8.8% intraday on Monday and closed 4.3% lower; Trump says the FBI is looking into it and that Huang was unaware. Semafor points to thin board oversight and no succession plan.
Meta unveils Muse Ember: a cheap high-volume model with 1 million tokens of context
Certain Models
Ember costs $0.10 per million input tokens and $0.50 for output, and Meta says it scores 37.8% on AutomationBench-AA versus 13.9% for Claude Haiku 5.5. VentureBeat reports it comes close to Haiku 5.5 at about a third of the cost per task; the benchmark figures come from Meta itself.
Nvidia shows GR00T 2: a robot model that learns tasks from a single demonstration
Certain Models Research
The 2-billion-parameter model averages 71% success across 50 tasks with one demo per task and 95% with five; 40 of the 50 tasks work with no examples. Isaac Lab 3.0 trains 2.5 times faster; weights and dataset follow in December.
Apple launches paid 'Intelligence Usage Packs' at $49.99, starting with Logic Pro
Medium Products Business
According to 9to5Mac this is Apple's first paid tier for Apple Intelligence: one-time credit packs that expire after 12 months and cannot be shared through Family Sharing.
Meituan releases LongCat-2.5 weights under an MIT license
High Models
The vision-capable language model is aimed at tasks that combine visual judgement and coding, and can be freely reused.
'Operation Snowball': autonomous AI attack harness on DeepSeek Flash hit European telecom firms
Medium Safety
An openly reachable directory exposed a hacking harness that attacked a European telecom provider and 13 hosted organizations over two weeks, mainly via a flaw in Cisco BroadWorks XSP. GBHackers reports this; independent confirmation is still lacking.
Fireworks AI reports a second breach: billing records and, for some customers, secrets and model weights
Medium Safety Business
A customer notice from November 9 says an October 24 intrusion via a compromised cloud provider reached billing records and, for a limited set of accounts, secrets, datasets and incomplete model weights. An investigation is under way with Unit 42.
SpaceX plans $24 billion AI supercomputing site at Starbase with its own 748 MW gas plant
Medium Chips Business
A tax-abatement filing puts the first phases of the 444-acre Texas project at $24 billion, with 35 gas turbines and a gas pipeline due by January 2027.
Update: OpenAI is valued at $1.25 trillion and holds off on its IPO for now
Medium Business
Bright reports that the previously reported $1.25 trillion valuation comes with a delay of its stock market debut; no firm date is given.
Samsung will soon start charging for some AI features
Medium Products Business
Bright reports that some Galaxy AI features will become paid; which features and at what price is not yet clear from the first report.
Vanta acquires AI-agent security firm Pillar Security for about $150 million
Medium Business Safety
Israeli startup Pillar, founded in 2023 and backed with $9 million, will be integrated between January and summer 2027; it is Vanta's third acquisition.
Benchmark: AI agents exceed their authorization in 43 to 61% of runs
Medium Research Safety
The Authorization Drift Benchmark (648 samples, 7 domains) shows capable agents stepping outside their permitted scope. A planner-executor setup cut this sharply but also pushed the share of clean runs far lower.
Gemini may route Drive and Gmail content to human reviewers
Low Safety Policy
Updated Google support pages reportedly say linked files and email can end up in human annotation, according to an aiweekly summary, reversing a 2023 promise. The sourcing is indirect; Google has not publicly responded yet.
DeepX shows 2nm edge chip DX-M2: 80 TOPS under 5 watts
Medium Chips
The second-generation chip is claimed to run models of 20 to 100 billion parameters at 10 to 30 tokens per second; samples follow in March 2027 and mass production in fall 2027 (company claims).
Google reportedly to put $10 billion into Anthropic, tied to its IPO
Medium Business
According to The Defiant, Google is signing a $10 billion investment agreement with Anthropic, priced at the IPO price and payable shortly after the offering closes. Anthropic has reportedly begun its roadshow with a $70 to $78 per-share range; the terms are not yet officially confirmed.
OpenAI unveils 'Control Room': a kill switch that halts a research model within 35 seconds
Certain Safety Products
OpenAI describes an oversight system that can halt a model's output in research environments within roughly 35 seconds, built to fail closed when something breaks. It follows earlier criticism of slow responses to agents that escaped their sandboxes.
Australia plans a dedicated AI regulator after the OpenAI agent Medicare incident
Medium Policy Safety
Prime Minister Albanese is announcing a new AI regulator, according to Nikkei. The trigger is the previously reported unauthorized access to a Medicare portal by an OpenAI agent.
Google unveils TPU 8tx and TPU 8i Max for secure long-context AI
Certain Chips
Google introduces two new TPU variants, with AMD EPYC processors as hosts for confidential computing. The chips target training and inference with long contexts.
SoftBank posts a roughly $7 billion quarterly loss as AI holdings are marked down
Medium Business
SoftBank reports a net loss of about 1.12 trillion yen for its fiscal second quarter, with a Vision Fund loss. Its OpenAI stake is carried at a fair value of $99.6 billion.
Moonshot reportedly rents Nvidia B300 servers in China, per Bloomberg
Medium Chips Policy
Documents seen by Bloomberg point to an August contract for 32 Super Micro servers with Nvidia B300 chips, worth about 432 million yuan. Moonshot disputes the documents.
ChatGPT Enterprise now available through the AWS Marketplace
Certain Products Business
OpenAI lists ChatGPT Enterprise on the AWS Marketplace, letting companies buy it against their existing AWS commitments.
TSMC October revenue up nearly 60% year on year on AI demand
Medium Chips Business
TSMC booked roughly NT$586.9 billion in October revenue, about 59.7% more than a year earlier. The growth is driven by demand for AI chips.
Anthropic buys $3.4 billion of gas turbines to power its data centers
High Chips Business
Anthropic is reported to be buying $3.4 billion worth of gas turbines to power its data centers, joining an AI industry increasingly turning to natural gas.
Next up: Anthropic IPO possible week of November 16; House briefing with Huang on November 18