Anthropic releases Claude Fable 5.2 at lower prices

Anthropic released Claude Fable 5.2 and Mythos 5.2 at lower prices, and Google made Gemini 4 Carbon broadly available. Also inside: Mercor's $750 million raise, Oura's Counsel Health deal, new robot models, and scrutiny of Nvidia's China chip sales.

AI lab releases

Anthropic releases Claude Fable 5.2 and Mythos 5.2 at lower prices

Anthropic released Claude Fable 5.2 for all users and Mythos 5.2 for vetted organizations, saying both improve coding, reasoning and vision. Both cost $8 per million input tokens, the small chunks of text that models process, down from $10 for the previous versions. Independent tester Artificial Analysis gave Fable 5.2 the top score so far, 65, but found it slow and costly per task.

Why it matters: Developers get a cheaper top Anthropic model per token, though heavy token use can still push up the cost of each task.

Sources: TechCrunch, The Verge, AnthropicAI, The Decoder, artificialanalysis.ai, The Information

Google makes Gemini 4 Carbon broadly available with agent swarm preview

Google released Gemini 4 Carbon to paying developers, businesses and AI Ultra subscribers after two weeks with a group of cybersecurity testers. It costs $2 per million input tokens at an introductory rate, which The Verge says will double at the end of November. Google also previewed /swarm in its Antigravity tool, which splits one request across a group of AI agents, programs that carry out tasks on their own.

Why it matters: Carbon gives buyers a much cheaper per-token option than rival frontier models, though real bills depend on how many tokens a task uses.

Sources: Google DeepMind, The Verge, The Decoder

DeepMind's Gemini Robotics 2.5 tested on Mercedes-Benz production lines

Google DeepMind released Gemini Robotics 2.5, a model that turns what a robot sees and is told into physical movements. DeepMind says an Apptronik Apollo 2 humanoid using the model sorted car parts with a 96% success rate in a Mercedes-Benz pilot. DeepMind also published ASIMOV-Jailbreak, an open test of whether robot models can be tricked into unsafe actions.

Why it matters: Factory robots that adapt to messy settings move closer to routine use, if the company's figures hold up beyond pilots.

Sources: Google DeepMind, The Decoder, Ars Technica

OpenAI lets businesses run verified agents inside ChatGPT

OpenAI launched a beta that lets eligible US online stores and service firms create Business Agents to answer customer questions in ChatGPT. Companies can claim and verify their listing through a Business Hub, which also shows how often ChatGPT mentions them. OpenAI says chats with these agents are labeled and kept separate, and that user conversations are not shared with businesses.

Why it matters: Shoppers may soon talk to brands directly inside ChatGPT instead of visiting their websites.

Sources: The Verge, The Decoder, VentureBeat

Acquisitions + deals

Mercor raises $750 million at a $20 billion valuation

Mercor, which supplies human experts to train and test AI models, raised $750 million in a Series D at a $20 billion valuation. Nvidia, Salesforce and Zoom joined the round, and no lead investor was named. The company plans to build custom tests that measure AI agents against each enterprise customer's own standards.

Why it matters: Firms deploying agents may get tailored ways to check them before trusting them with real work.

Sources: SiliconANGLE, Reuters

Oura agrees to buy medical AI startup Counsel Health for $625M

Smart ring maker Oura is acquiring Counsel Health, whose app pairs a medical chatbot with access to licensed doctors, for $625 million. Counsel's service has been part of the Oura app since June, and its 70 employees will join Oura. The deal comes weeks after Oura postponed its planned IPO indefinitely.

Why it matters: Oura ring owners could get medical advice and doctor access tied to their own health data in one app.

Sources: TechCrunch, The Verge

Crunchbase: 34 new unicorns in October push board past 1,900

Crunchbase News reports 34 companies joined its Unicorn Board, which tracks private startups valued at $1 billion or more, in October. Crunchbase News says half of the new unicorns are under four years old, and nine got there through seed or Series A rounds. According to Crunchbase News, the most valuable newcomers included AI model maker TypeSafe AI and agent company Manus AI.

Why it matters: Young AI companies are reaching billion-dollar valuations quickly, with investor money flowing into very early rounds.

Sources: Crunchbase News

OpenAI acquires AI travel agent startup Odessia

OpenAI has bought Odessia, an AI travel planning startup, according to its founder Francis Davidson. Davidson, who earlier founded rental company Sonder, is joining OpenAI to work on travel, and terms were not disclosed. Odessia's tool has already been shut down, and OpenAI has not announced any travel products.

Why it matters: Trip planning inside ChatGPT may get more attention, given how often users already ask it about travel.

Sources: TechCrunch, The Information

More funding

• SiliconANGLE reports Path Robotics, which uses AI to run industrial welding robots, raised a $300 million Series E led by Georgian. Sources: SiliconANGLE
• SiliconANGLE reports Dreeze launched with $33 million in seed funding for software that restores whole cloud applications after outages. Sources: SiliconANGLE
• SiliconANGLE reports Daylight Security, which runs security operations with AI agents, raised an $80 million Series B led by Goldman Sachs Alternatives. Sources: SiliconANGLE
• SiliconANGLE reports Staera Semiconductor launched with $4.6 million in pre-seed funding to build optical links between chiplets. Sources: SiliconANGLE
• SiliconANGLE reports PuppyGraph, whose engine lets AI agents query company data as a graph, raised a $10 million Series A. Sources: SiliconANGLE
• SiliconANGLE reports Lumbra raised a $25 million Series A for a platform that runs AI agents in classified environments. Sources: SiliconANGLE
• SiliconANGLE reports Bay AI launched with $50 million across seed and Series A rounds to check AI agent actions before they run. Sources: SiliconANGLE

MCP + agent tooling

Red Hat and Nvidia target AI agent operations at KubeCon

Red Hat released OpenShift 5, the first major update in seven years to its Kubernetes platform for running apps across many servers. The update adds an AI agent that investigates operational problems and proposes fixes, with administrators deciding whether to apply them. Nvidia is donating OpenShell, its open source software for securing and controlling AI agents, to the Cloud Native Computing Foundation.

Why it matters: IT teams get more help from AI on routine fixes, along with shared tools to limit what agents can touch.

Sources: SiliconANGLE, NVIDIA Blog

AWS launches web search API for Bedrock AgentCore agents

AWS made a Web Search API available in Amazon Bedrock AgentCore, giving AI agents current web results through a single call. AWS says queries are served from a web index Amazon runs and are not sent to an outside search engine. Each result includes a title, link, publication date and relevant extract, and requests can filter by domain or date.

Why it matters: Teams building agents on AWS can add live web lookups without handing their queries to a third party.

Sources: Amazon Web Services

Browser Use 1.0 rewrites its browser agent around direct Chrome control

Browser Use released version 1.0 of its browser automation library, deleting about 80% of the code in a rewrite. Its agent now writes Python that drives Chrome directly through the Chrome DevTools Protocol, instead of picking one action per step. The maker claims, from its own tests, that the same models score up to 33 points higher and run 2.7 to 5.2 times cheaper.

Why it matters: Developers relying on the old version will need to plan a migration to get the claimed speed and cost gains.

Sources: Browser Use on GitHub

Simon Willison uses Claude Code to track down a mystery beep

Developer Simon Willison had Claude Code build Beep Locator, which uses microphones on several devices to pinpoint where a sound comes from. The tool placed an intermittent beep in a room full of boxes to within about 40 centimeters. The source turned out to be an old carbon monoxide detector that needed new batteries.

Why it matters: Coding assistants are letting individuals build one-off tools for small household problems.

Sources: Simon Willison, GIGAZINE

SDKs + APIs

LiteLLM 1.109.0 adds chat interface and mid-stream fallbacks

LiteLLM released version 1.109.0, a large update spanning 982 pull requests. It adds a chat interface to the admin dashboard and lets a streaming answer cut off midway continue on a backup model. Five changes require action from users, including the removal of the disk cache backend.

Why it matters: Teams running LiteLLM should review the breaking changes before upgrading.

Sources: LiteLLM on GitHub

GitHub Copilot adds Gemini 4 Carbon for paid plans

GitHub made Gemini 4 Carbon generally available in Copilot for Pro+, Max, Business and Enterprise users. It appears in the model picker across VS Code, JetBrains IDEs, Xcode, the Copilot CLI and other tools, with a gradual rollout. Use of the model is charged at the provider's list price under usage-based billing.

Why it matters: Copilot subscribers can try Google's newest model without leaving their usual editor.

Sources: The GitHub Blog

Research

Nvidia unveils GR00T 2 robot model that learns from human video

Nvidia introduced GR00T 2, an open model that helps robots learn delicate hand tasks, at the CoRL robotics conference. The Decoder reports the model predicts a short video of what happens next, which lets it transfer human movements to robot arms. Nvidia says it reached 71% average success on 50 tasks with one demonstration each, rising to 95% with five.

Why it matters: Teaching robots new jobs could require far less costly remote control by human operators.

Sources: NVIDIA Blog, The Decoder

Startup PSI uses Anthropic model to prove quantum computing conjecture

Physical Superintelligence says its research system, running an unreleased Anthropic model, proved the Aaronson-Ambainis conjecture, open since 2008. The proof was checked in Lean, software that verifies each step of a mathematical argument. Scott Aaronson, who co-posed the problem, wrote that he believes the proof is correct.

Why it matters: AI systems are starting to settle long-open questions in theory, leaving experts to check and interpret the results.

Sources: Import AI, TechCrunch

OpenAI and Upwork study tracks AI's effect on freelance demand

OpenAI and Upwork compared freelance job postings with ChatGPT use to see how AI is changing demand for freelance work. Many tasks common in ChatGPT lost share on Upwork, a pattern found in 43 of 63 work categories. AI-related projects had larger fixed budgets and more hours, and often combined tasks that rarely appeared together before.

Why it matters: Freelancers who can combine AI with other skills may find bigger projects, according to these early findings.

Sources: OpenAI

Other notable news

Huang affair puts Nvidia's push for China chip sales under scrutiny

Jensen Huang's affair with an alleged Chinese spy is drawing new attention to his campaign to ease US limits on chip sales to China. TechCrunch reports Nvidia earned $7.9 billion from China-based customers in the quarter ended July 26, about 8% of revenue. The Verge, citing The New York Times, says Megaspeed's sister company spent at least $4.8 billion on Nvidia chips and technology.

Why it matters: Any loss of trust in Washington could make it harder for Nvidia to win approval for future sales to China.

Sources: TechCrunch, CNBC, The Verge

Axios: Commerce Secretary Lutnick got classified warning on Huang affair

Axios reports Commerce Secretary Howard Lutnick received a highly classified memo about Jensen Huang's relationship with suspected spy Alice Huang. Lutnick runs the agency meant to keep sensitive US technology from adversaries, which cleared Nvidia chip sales to Chinese companies. Axios says Nvidia has lost about $1 trillion in market value since the relationship was first reported, partly because of election results.

Why it matters: Questions over who knew what, and when, could put the government's own chip export approvals under review.

Sources: Axios

AMD launches Epyc 6 server chips aimed at AI agent workloads

AMD launched its Epyc 6 server processors, known as Venice, with ten chips built on a new Zen 6 core design. The lineup ranges from 96-core high-speed chips to 256-core chips meant to run many agents at once. AMD claims, based on its own figures, that the chips run more AI agents per watt than Nvidia's Vera processor.

Why it matters: Data centers running agent tools now have another choice of server chips.

Sources: SiliconANGLE, Neowin

Cyanogen co-founder unveils AI-focused Linux desktop OS

Kirt McMaster, who co-founded Cyanogen, unveiled Kora, a Linux-based desktop operating system that builds AI models and agents into the computer. It is planned as a free release in 2027 for PCs with Intel, AMD or Qualcomm Snapdragon chips. McMaster said his team of about 50 engineers used AI coding tools that sped up development at least fivefold.

Why it matters: Heavy AI users may get a Windows alternative designed so agents can work across apps and files.

Sources: TechCrunch, Wired

Thanks for reading, and we will be back tomorrow with more.

---

About ScrutifAI: ScrutifAI is an independent third-party agency that tests software for AI agents, building a single access point where agents and people can access and use tested tools without learning each one individually, at a fraction of the cost of buying them directly.