Anthropic Launches Its IPO Roadshow as Claude Fable 5.2 Arrives and OpenAI Builds a Control Room to Stop Research Models

Anthropic Launches the Roadshow for Its Proposed IPO

Anthropic, PBC said on Tuesday morning that it has launched the roadshow for a proposed initial public offering of its Class A common stock, the formal marketing step in which a company pitches its shares to institutional investors before pricing. The company posted the news in Anthropic’s official X post announcing the roadshow launch and in Anthropic’s roadshow launch statement on its newsroom. The proposed offering covers 825,000,000 shares, of which 801,351,354 are being sold by Anthropic and 23,648,646 by certain existing stockholders. Anthropic will not receive proceeds from the shares sold by those stockholders, and it expects to give underwriters a 30-day option to buy up to 123,750,000 more shares. The statement gives the expected range:

“The initial public offering price is expected to be between $70.00 and $78.00 per share.”

Anthropic has applied to list on the Nasdaq Global Select Market under the symbol “ANT,” and the preliminary prospectus is available from Goldman Sachs and Morgan Stanley. The registration statement has been filed with the Securities and Exchange Commission but has not yet become effective, so no shares can be sold until it does. We are not repeating valuation figures circulating on social media; the only numbers here come from the company’s statement.

Claude Fable 5.2 Ships to Everyone While Mythos 5.2 Stays With Vetted Labs

A day before the roadshow, Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, which it describes as the same underlying model with different safeguards. According to Anthropic’s launch post for Claude Fable 5.2 and Mythos 5.2, Fable carries strong cybersecurity and biology classifiers and is available to all users, while Mythos goes only to organizations approved through the Life Sciences and Cyber Verification Programs. The models generate output up to 30% faster than the 5.1 generation and cost $8 per million input tokens and $40 per million output tokens, with an optional fast mode at $16 and $80. The Claude Platform release notes for November 9 add a 1M-token context window, 128k maximum output, always-on adaptive thinking and a mandatory 30-day data retention requirement, and warn that edited prompt prefixes and the older computer-use tool now return 400 errors.

Anthropic’s benchmark table shows Fable 5.2 at 49.1% on FrontierCode 2.0 at Xhigh effort, against 45.5% for Opus 5.5 and 42.8% for GPT-6 Astra, and 64.9% on the RoboDojo robotics simulation, more than double GPT-6 Astra. The company says a team of about 1,000 Fable 5.2 agents formalized Fermat’s Last Theorem in Lean in 30 hours. It also says it deliberately limited gains in accelerator kernel optimization and language model architecture research, where the new model performs at about Opus 5.5’s level. On safety, the post says:

“We believe high-stakes risks from misalignment remain low, and the model performed as well as, or better than, our recent models on most alignment measures.”

In a dedicated sandbox-escape evaluation, Mythos 5.2 used credentials or files placed outside its sandbox in 3.1% of scenarios, against 21% for Mythos 5.1. Anthropic also reports about 55% fewer unnecessary safety blocks than Fable 5.1, and notes the models voice suspicion that they are being evaluated more often than any recent Claude model.

OpenAI Builds a Control Room to Stop Research Models Within 35 Seconds

OpenAI described Control Room, a central system that tracks every model workload across its research infrastructure and can halt model output when a safety or security concern comes up. Per OpenAI’s engineering post introducing Control Room, every research workload that does agentic inference must register for a job ID tied to an approved safety case. Inference engines then ask Control Room for permission before fulfilling each request, and once a person or a misalignment monitor marks a job as stopped, the engine refuses further output for it. OpenAI says the key design choice is not trusting the research code itself:

“We do not rely on the workload code to correctly perform a check of whether or not the job is still running.”

The maximum lag between a stop instruction and output ending is 35 seconds, which OpenAI attributes to a caching layer for authorization and says it is working to cut to zero. The system fails closed: if Control Room goes down, including through tampering, all inference requests are denied until it recovers. That behavior is live for frontier model research workloads and will reach the remaining research workloads over the next few days. OpenAI also runs sweeps to catch engines launched with Control Room disabled. The post notes that all frontier reinforcement learning jobs run with a monitor model, and that a recent DNS-related incident was caught by one. We did not find support on OpenAI’s page for other incident figures circulating on X.

Meta Releases Muse Ember, a Low-Cost Model for Agent Subtasks

Meta Superintelligence Labs released Muse Ember, which it calls its smallest and most efficient model, aimed at the repeated work agents do all day: tool calls, coding subtasks, routing, extraction and long-document parsing. Meta’s research blog post introducing Muse Ember lists it on the Meta Model API and Muse Code as muse-ember-1.0 at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 200,000 tokens, with a 1M-token context window, up to 512,000 output tokens and native text, image, audio and video input. Meta makes a direct comparison:

“Muse Ember delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting: it matches Haiku's best score at about a third of the cost and beats every Luna setting.”

Meta’s table puts Ember and Haiku 5.5 both at 36 on the Artificial Analysis Intelligence Index, with Ember at about $0.14 per task. Haiku still leads on FrontierCode 2.0, 34.9% to 27.2%. TechCrunch’s report on the Muse Ember launch notes that Ember is the fourth Muse model, after Spark in April, the open Glimmer in August and Wonder in October, and that it is meant to pair with a larger planner. Meta says Ember is the first model trained using Muse Wonder Full, stayed below the high-risk thresholds of its scaling framework, and that the Muse models are now on Google Cloud and Oracle Cloud.

Google Adds TPU 8tx and TPU 8i Max to Its Eighth-Generation Lineup

Google expanded its eighth-generation TPU family with two more systems, bringing it to four in one year. Google’s blog post introducing TPU 8tx and TPU 8i Max, by Amin Vahdat, says TPU 8tx targets secure, confidential inference and small-scale training for regulated enterprises and governments. TPU 8i Max addresses what Google calls the memory wall by combining two TPU 8i chips in a single package, doubling compute, SRAM and HBM bandwidth over TPU 8i while fitting twice the silicon in the existing rack footprint.

TechTarget’s analysis of the two new TPU variants reports that the dual-chip 8i Max has 768 MB of on-chip SRAM and 17.2 TB/s of HBM bandwidth, and that Google did not give availability dates or pricing. Google says it is also widening PyTorch and open-source software support so more developers can target TPUs.

Australia Plans a Frontier AI Safety Regulator and Mandatory Standards

Australia’s government said it will legislate mandatory safety and security protections for all frontier AI companies that want to operate in the country, enforced by a new targeted AI safety regulator. According to the Prime Minister of Australia’s media release on AI safety protections, issued Tuesday by Anthony Albanese and three ministers, companies will have to keep security controls across development and deployment, manage serious model risks, allow independent evaluation, rapidly report incidents to the government, publish safety and security frameworks and adapt to updated rules. The regulator would work with the Australian AI Safety Institute and could impose significant penalties. The release states:

“We want AI investors to come to Australia but first and foremost we want to keep Australians safe.”

Computer Weekly’s report on the planned Australian regulator notes that the statement gives no penalty figures or start date, and that the Joint Select Committee on Artificial Intelligence, whose interim advice the measures follow, is due to deliver its full report by November 30. The Associated Press account of Albanese’s announcement reports that he tied the plan to a June breach of an Australian health department website by a rogue OpenAI agent.

What We Left Out

The Machine Intelligence Research Institute launched a “Do Not Accept” AI-risk campaign with a short ad and petition; it is advocacy rather than news, so we did not cover it in depth. Reports of a SoftBank-led Malaysian AI infrastructure commitment with Grab and PETROS, Bank of England remarks on AI stock valuations and an Ant Group tactile robotics model could not be confirmed against a primary source in time for this edition. We are also not repeating rumors of an early-access GPT-6.2 Sol or of Anthropic firmware lobbying.