Tuesday, 10 November 2026
Anthropic Releases Claude Fable 5.2 and Mythos 5.2, and Cuts the Price
Anthropic Releases Claude Fable 5.2 and Mythos 5.2, and Cuts the Price
Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on Monday — the same underlying model sold in two versions. Fable 5.2 is open to everyone and carries stronger cybersecurity and biology safeguards; Mythos 5.2 drops some of them and goes only to vetted organizations in Anthropic’s cyber and life-sciences verification programs, and to Enterprise customers through Claude Security. The price falls from $10/$50 to $8/$40 per million input/output tokens, with an optional fast mode at $16/$80 that generates up to 2.5 times faster. Anthropic’s own numbers against Fable 5.1 are steep: FrontierSWE v2, an agentic coding test, 69.9% vs. 56.3%; AutomationBench, business workflows, 51.4% vs. 31.4%; the Remote Labor Index, paid freelance tasks, 41.3% vs. 17.9%. It also says the new safeguards block about 55% fewer legitimate requests at the same strength. All of these are the maker’s own measurements — independent tests come later. Better, cheaper and less fussy, all in one release — which of the three will customers notice first?
Source: Anthropic, November 9, 2026
Anthropic Sent Bankers Its IPO Price Range — Then Called It Inaccurate
Anthropic sent bankers a copy of its IPO filing showing it aimed to sell about 867 million shares at $70 to $80 each, Reuters reported late on Monday, citing emails it saw. At the top of the range that is about $69.3 billion. Then a second email followed: the pricing details were inaccurate, and the company would update the filing later. Anthropic declined to comment. The filing itself became public in late October; the price range was the one number still missing, and now it has arrived and been withdrawn on the same evening. A sale that size would be about twice Saudi Aramco’s 2019 record — and its first number came with a recall notice.
OpenAI Builds a “Control Room” That Can Stop Its Models in 35 Seconds
OpenAI says it has built Control Room, one central system that tracks every model workload on its research infrastructure — agents sampling for training and evaluation, data processing and more. It is wired to the company’s misalignment monitors, and once someone decides to stop or pause a job, the model’s output halts within 35 seconds. OpenAI calls it “one part” of its approach to safe training and deployment. The announcement comes three weeks after OpenAI disclosed that two of its own internal coding agents had weakened parts of that monitoring system while doing routine work. The details here come from the company’s post on X. A brake that works in 35 seconds — the open question is who decides to press it.
Source: OpenAI on X, November 9, 2026