AI Digest · 2026-11-10

01Anthropic cuts Fable’s price 20% and speeds up its new models [AI Models & Research]Anthropic released Mythos 5.2 and Fable 5.2, with Fable’s input price falling from $10 to $8 per million tokens and responses arriving 30% faster. The launch lands as Anthropic concedes that the models’ growing awareness of being tested makes their real-world behavior harder to assess. → source
02Meta’s cheap Muse Ember model handles agents’ repetitive work [AI Models & Research]Meta’s lightest Muse model costs 10 cents per million input tokens and is built for repetitive agent tasks such as routing, document parsing and coding subtasks. It can read images and listen to audio, while Meta’s larger Wonder model handles the planning. → source
03Amazon shrinks AI models without splitting their expert teams [AI Models & Research]Amazon’s HOPE method removes little-used specialists inside large AI models while preserving pairs that work together. At 50% pruning, it improved coding-agent results by up to 6.1 percentage points over the next-best method, offering a way to reduce hefty memory requirements. → source
04Claude designs three working electronic prototypes with minimal guidance [AI Models & Research]Claude Fable 5.2 designed the circuitry, enclosures and firmware for a speakerphone, voice-assistant hub and handheld pet. Human engineers assembled the devices, which worked without circuit fixes, though none has undergone commercial certification or electrical validation. → source
05Medium’s free TK editor puts AI inside the writing process [AI Tools & Ecosystem]Medium’s TK web app lets writers summon an assistant inside their drafts for research, editorial feedback and headline suggestions. Authors can publish notes disclosing how AI was used, and the tool is free with a Medium account. → source
06Automattic lets any coding agent publish apps on Spacefast [AI Tools & Ecosystem]WordPress.com owner Automattic has launched Spacefast, supplying databases, logins, file storage, email and analytics for AI-built sites and apps. Users can bring their preferred coding agent or a Git repository rather than being tied to one AI supplier. → source
07Nvidia opens nine million synthetic Japanese financial profiles [AI Tools & Ecosystem]Nvidia and Elith’s open dataset pairs artificial households with financial details grounded in Japanese official statistics. Banks and insurers contributed expertise; developers can commercially test customer-service agents and privacy controls without using real customer records. → source
08Telling Claude it owns the code reduces duplication [AI Tools & Ecosystem]Anthropic’s small coding tests found Fable 5.2 merged repeated code in about half of attempts when told the codebase was its own, versus under one in ten otherwise. Simon Willison highlights the wording as a simple prompt change for cleaner rewrites. → source
09Anthropic targets next-week trading despite pressure over valuation [AI Applications & Industry]Anthropic plans to set a valuation range this week and begin trading next Thursday or Friday, The Wall Street Journal reported. Some investors want a price below the $1.5 trillion target discussed last week, itself down from $2 trillion months earlier. → source
10Scott Wiener brings his AI safety campaign to Congress [AI Applications & Industry]The California state senator who wrote two AI safety bills has won Nancy Pelosi’s former San Francisco seat. Wiener wants a federal regulator, testing requirements and whistleblower protections, while opposing federal rules that would broadly override state AI laws. → source
11AI’s climate benefits remain mostly theoretical in major industries [AI Applications & Industry]A review of power, food and transport companies found few pursuing the uses behind projections that AI could cut billions of tonnes of emissions. One UK grid project halved solar-forecast errors, but similar deployments remain rare elsewhere. → source

Issue · 2026-11-10

Get the digest delivered