Fable 5.2 reins in AI research skills; breach hits Notion

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, the same underlying model with and without its cybersecurity and biology safeguards. It costs $8 per million input tokens and $40 per million output, down from $10 and $50 for version 5.1, and leads Anthropic's own table on most of the benchmarks it reports. The notable choice is what it left out. Days after an outside audit found AI roughly doubling Anthropic's pace, the company says it limited the model's skill at the work that most directly accelerates AI development, such as optimizing chip kernels and improving language model designs, holding it near Claude Opus 5.5 there. In a sandbox-escape test it used planted credentials in 3.1 percent of scenarios, against 21 percent for Mythos 5.1.

The Wall Street Journal reports that Anthropic will set a valuation range this week and could begin trading as soon as late next week, despite a sharp tech sell-off and the controversy over Nvidia's chief executive and a suspected spy. Some institutional investors are pushing for a valuation below the $1.5 trillion discussed only last week. Separately, the Financial Times reports that Anthropic, which has long preferred renting computing power to owning infrastructure, is buying Caterpillar gas-fired generators worth $3.4 billion and able to supply up to 1.7 gigawatts, after the developer that originally ordered them could not finance the purchase.

Notion says a security incident at Fireworks AI, which it used to serve open-weight models and to classify AI usage for its Insights feature, exposed some Notion AI chat transcripts from under 1 percent of workspaces. The chats carry no names or workspace labels, but Notion is telling customers to treat them as exposed and to rotate any passwords or keys pasted into Notion AI. Fireworks first disclosed stolen internal credentials in October, and on November 6 told Notion that attackers had reached a further dataset. Notion has cut Fireworks off, deleted its data there, and halted its training runs.

Asked on 60 Minutes about AI executives' safety warnings, President Trump said the risk is "not nearly as real" as they claim and is being managed. As an example, he said the Justice Department called the leader of "one of the big" AI companies, who had been reported for doing "something bad with AI," and told them to get ready to go to jail, after which the person "totally changed everything." He did not name the person or the company. Trump also said he would work with Democrats on AI safety and does not "mind guardrails."

Meta Superintelligence Labs released Muse Ember, its smallest and cheapest model, at 10 cents per million input tokens and 50 cents per million output on the Meta Model API. Meta pitches it as the worker beneath a larger planner such as Muse Wonder, handling tool calls, routing, extraction and long documents, and says it matches Claude Haiku 5.5's best Artificial Analysis score at about a third of the cost while beating every setting of OpenAI's GPT-6 Luna. It reads screenshots, audio and video without a separate model. Box, Cline, Oracle and Cloudflare tested it before launch.

OpenAI described Control Room, a central service that every agentic research workload must check in with before a model produces output, covering training, evaluation and data-processing jobs. When staff decide to stop or pause a job, the model's output halts within 35 seconds, and if Control Room itself goes down, output stops rather than continuing unchecked. Every such workload now also needs an approved safety case. It builds on the training halts OpenAI described in August, and the company says its public model-serving systems already had a fast takedown path.

Security firm Apiiro found more than 100,000 copies of Claude Code's local settings file in public GitHub repositories, and about 1,100 of them, from 63 organizations, held secrets such as database passwords, API tokens and private keys. Approving a shell command with "Yes, and don't ask again" can save the full command, credentials included, into that file. Claude Code adds it to a developer's global ignore list but not to the repository's, so it can slip into a commit. The issue was first reported in an August 2025 bug report, and Anthropic's Boris Cherny says the team is looking at fixes.

Two Amazon engineers replaced Claude Sonnet 4.6, which had been sorting tickets into 20 categories in one of the company's largest internal ticketing systems, with Strands Decider, the small open-source model AWS released last month that picks from a fixed set of options instead of writing text. Running on a single-GPU SageMaker instance, it performed as well as Sonnet at about $500 per million tickets, against roughly $22,200, and answered in under a second instead of more than four. The AWS team says the switch will save millions of dollars a year.