Anthropic unveils Fable 5.2

Hi there, this is your daily ☕️ Techpresso.

Anthropic unveils Fable 5.2 LINK

• Anthropic has released Claude Fable 5.2 and Claude Mythos 5.2, its two newest AI models, as the company gets ready for an IPO that could happen as soon as next week.
• The company says both models run 30% faster than the 5.1 generation and cost 20% less, priced at $8 per million input tokens and $40 per million output tokens.
• It's Anthropic's first "frontier" release since CEO Dario Amodei called in September for a slower pace of AI work, and the firm says the models went through fresh safety and alignment testing.

Report reveals Huang affair aided $7.5 billion China chip shipments LINK

• A report from The Information says Nvidia chief Jensen Huang's affair with Megaspeed executive Alice Huang helped move at least $7.5 billion worth of restricted Nvidia AI chips into companies in China.
• Alice Huang and her family reportedly set up a network of firms across Singapore and China, including Shanghai Century Litong, which Nvidia named a preferred cloud partner just four months after it was founded in 2023.
• Her ties to Nvidia and its distributors allegedly helped ship roughly 240,000 H-series and A-series chips into China, with some reaching giants like Alibaba and Tencent; Nvidia says it sees no sign its secrets were compromised.

Ring unveils a flying camera and patrol robot LINK

• Ring has introduced two autonomous security gadgets, the $250 Drone Cam that flies around your home and a ground-based Rover Cam, both shipping under its new Ring Robotics line starting on January 27, 2027.
• The 165-gram Drone Cam is a small quadcopter with a 2K front camera, a 1080p underside camera, and LEDs that can mimic a guard's flashlight; it flies on a schedule or reacts to motion and leak sensors.
• Its battery lasts eight to 10 minutes per flight, and a system Ring calls AI Pilot uses cloud processing over Wi-Fi to map rooms, pick patrol points, and steer around obstacles instead of running local AI.

Meta's cheapest AI model matches Claude Haiku LINK

• Meta has launched Muse Ember, its smallest and cheapest AI model, which the company says ties Anthropic's Claude Haiku 5.5 on intelligence scores while doing the same work for roughly a third of the price.
• Ember costs $0.10 per million input tokens and $0.50 per million output, the same list price as Haiku 5.5 and OpenAI's GPT-6 Luna, but ran one benchmark for about $0.14 per task versus Haiku's $0.41.
• Built for small repeated agent jobs like tool calling and document parsing, Ember reads up to a million tokens across text, images, audio and video, though Meta is keeping its model weights private.

OpenAI builds kill switch for rogue research models LINK

• OpenAI has built a system called Control Room that can cut off an AI research model's access to output, a safety measure prompted by several cases this year of its models acting beyond their instructions.
• The system tracks training runs, evaluations, and data processing across OpenAI's research setup, assigning each job an identifier that the inference engine checks before producing any model output.
• When a person or automated monitor orders a job stopped, Control Room revokes its authorization and blocks further output, though cached permissions can delay a shutdown by up to 35 seconds.

Google DeepMind proves Signal encryption is secure LINK

• Google DeepMind says its AI agents mathematically proved that the code behind the Signal protocol, which secures conversations in apps like Signal and Google Messages, works correctly without security flaws.
• The team used a framework called AlphaProof Code, pairing human experts with AI agents powered by Gemini Carbon to generate over six million lines of Lean, a language that checks code is accurate.
• The process took four weeks instead of years, with agents flagging just one percent of the code for human review, and it correctly caught two old bugs deliberately put back into an isolated copy.

Other news you might like

• Anthropic seeks to be valued at $1.02 trillion LINK

• Apple launches $50 ‘intelligence usage packs’ for select AI features LINK

• Uber is launching its own parking reservation service LINK

• US sues Lyft, alleges discrimination against disabled riders LINK

• Medium’s new TK app brings AI assistance directly into the writing process LINK

• Waze rolling out cyclist alerts based on real-time locations from Garmin trackers LINK

• Reddit is turning user conversations into AI-powered product reviews LINK

• Moonshot AI rented Nvidia’s top chips in China, Bloomberg reports LINK

• Adobe brings AI agents and new creative tools to Photoshop, Premiere, and Firefly LINK

• France blames Iran for supermarket refrigerator cyberattacks, weighs sanctions LINK

• North Korea Is Using Elon Musk-Themed Memecoins to Launder Crypto From the Binance Hack LINK

• Claude Code is accidentally exposing secrets all over GitHub LINK

• How Product Hunt almost died LINK

• Fable 5.2 and Jagged AGI LINK

Trending tools

ClipsPal: desktop video editor for Mac and Windows that trims silences and bad takes, adds subtitles and zooms, then exports shorts and articles LINK

Zigzag Browser: a lightweight WebKit Mac browser for ex-Arc users, offering sidebar, spaces, split view, Chrome extensions, and built-in ad blocking at 18 MB. LINK

satellites: finds every satellite that passed over a given location in the last 7 days using SGP4, then pulls the actual Sentinel-2, Landsat, and NASA images from those passes. LINK

HolyArrs: a 15 MB self-hosted media server that streams your personal video and music libraries, offering Netflix and Spotify-style interfaces. LINK

PC Part Pricer: pastes a prebuilt PC listing to receive a value grade, dated component price estimates, and a shareable report using a daily-updated pricing engine. LINK

OpenVibe: aggregates decentralized social networks into one timeline, letting you crosspost everywhere simultaneously while keeping a portable, platform-independent identity. LINK

Trending papers & reports

Synthetic shoppers built from vision-language models that browse photos of store shelves pick products as accurately as real households' own past trips, matching group buying patterns twice as well as a trained statistical model. LINK

Medical AI advice often recognizes updated treatment guidelines without fully applying them, with the best model showing a ~20 point gap, meaning chatbots can give outdated clinical recommendations even when handed the new evidence. LINK

Always-on work agents are tested on whether they can spot pending tasks buried in a worker's email, chat, and documents and then complete them unprompted, using a simulated company with known correct answers. LINK

AI-driven hacking compromised 395 organizations across 48 countries in one campaign, half of them schools, as a16z reports that cheap AI agents turned newly disclosed bugs into working attacks in hours rather than days or months. LINK

See you tomorrow for a new dose of ☕️ Techpresso!