Anthropic releases Claude Fable 5.2 at lower prices

Anthropic released Claude Fable 5.2 and Mythos 5.2 at lower prices, and Fable 5.2 has already reached Amazon Bedrock, GitHub Copilot and Claude Code. You will also find Google's new TPU systems, Shield AI's deal for Echodyne, Nvidia's Korean investments, and questions about Nvidia's leadership.

AI lab releases

Anthropic releases Claude Fable 5.2 and Mythos 5.2 at lower prices

Anthropic launched Claude Fable 5.2 for all users. Mythos 5.2 uses the same underlying model with fewer safeguards and is limited to vetted organizations. Both cost $8 per million input tokens, the small chunks of text that AI use is billed in. Anthropic says both run up to 30% faster than the previous versions. Artificial Analysis, an independent testing site, ranked Fable 5.2 first on its index, but at the top setting it costs about $26.50 per task.

Why it matters: Developers get a lower per-token price on a top model, but heavy token use can still push the real cost of each job up.

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

Google adds two new TPU systems for secure and memory-heavy AI

Google added two systems to its eighth generation of TPUs, its in-house AI chips: the TPU 8tx and the TPU 8i Max. Google says the 8tx keeps data and model weights protected even during processing, and it is aimed at governments and regulated firms. The 8i Max puts two chips in one package, which Google says doubles compute and memory within the same rack space.

Why it matters: Organizations with sensitive data get another hardware choice as Google sells more of its chips to outside customers.

Sources: The Decoder, SiliconANGLE, The Information

Apple starts selling $49.99 packs for extra AI use in Logic Pro

Apple now offers a $49.99 one-time purchase in Logic Pro that lets users keep using its AI assistant after reaching usage limits. The limits cover Apple Intelligence features that depend on cloud models from OpenAI or Google Gemini. Unused credit rolls over each month but expires after 12 months, and the packs are expected in other Apple creative apps this month.

Why it matters: Heavy users of Apple's AI features may soon face extra charges across more of its apps.

Sources: The Verge, 9to5Mac

Nvidia's Nemotron 3.5 Super VL arrives on Amazon SageMaker JumpStart

AWS says Nvidia's Nemotron 3.5 Super VL can now be deployed through Amazon SageMaker JumpStart in a few steps. The model takes images, video and text and replies in text, using 120 billion parameters, the values a model learns, with 12 billion active at once. Nvidia designed it for reading documents and helping agents understand and operate software screens.

Why it matters: AWS customers can test a large model that handles images and video without setting up their own servers.

Sources: Amazon Web Services

Acquisitions + deals

Hardware-testing startup Nominal raises $240M at $3B valuation

Nominal, which makes AI software for testing hardware, raised $240 million in a Series C round led by Lightspeed Venture Partners and Insight Partners. The funding values the company at $3 billion, three times its valuation in March. Nominal says over 80 commercial and government groups use its platform, including five of the seven largest defense contractors.

Why it matters: Investors are putting large sums into AI tools for physical engineering, not only software.

Sources: SiliconANGLE, TechCrunch

Shield AI agrees to buy radar maker Echodyne for up to $5.3B

Defense tech company Shield AI agreed to buy Echodyne, which builds compact radar that tracks drones, in a deal worth up to $5.3 billion. Shield will pay $4.8 billion at closing plus up to $500 million tied to targets, and is raising $2.5 billion to help. Separately, Docusign bought Suplari, a Seattle AI spending-analysis startup that Microsoft bought in 2021 and later spun back out.

Why it matters: Deals like these suggest buyers are again paying for AI and defense startups.

Sources: GeekWire, SiliconANGLE

TechCrunch AI reports Anthropic is buying $3.4B of gas turbines

TechCrunch AI reports, citing the Financial Times, that Anthropic is spending $3.4 billion on Caterpillar natural gas turbines. TechCrunch AI says the generators were first ordered by supplier OnePWR and can produce up to 1.7 gigawatts of power. TechCrunch AI notes Anthropic has usually rented computing power instead of buying physical infrastructure.

Why it matters: Electricity shortages are pushing AI labs to buy their own power sources.

Sources: TechCrunch

Nvidia plans investments in four Korean AI startups

Nvidia announced an investment in Korean startup Point2, which builds radio-based links that move data between chips and racks. It also plans to invest in Motif Technologies, RLWRLD and Upstage, which work on AI models, robotics and document tools. The news came at a forum Nvidia held with Korea's Ministry of Science and ICT, attended by more than 20 startups.

Why it matters: Korean startups gain a major backer as Nvidia invests across the AI supply chain.

Sources: NVIDIA Blog

MCP + agent tooling

Browser Use releases version 1.0, a near-total rewrite

Browser Use, a library that lets AI agents control web browsers, shipped version 1.0 after deleting about 80% of its code. Its agent now writes Python that drives Chrome directly, and it defaults to the company's cloud browser and model service. The makers claim up to 33 points higher scores with the same models, at 2.7 to 5.2 times lower cost.

Why it matters: Existing users must check migration notes, since default settings now route work through a paid cloud service.

Sources: Browser Use on GitHub

Stripe built company-wide AI assistant Kai in a week, LangChain says

LangChain says Stripe built Kai, an internal AI assistant for all employees, in one week using its open-source Deep Agents framework. Kai connects to Stripe's data warehouse, Slack and Google tools to produce documents, reports and dashboards. More than 100 teams have added over 1,000 skills, packaged know-how the assistant can draw on.

Why it matters: Non-engineers at large firms are getting agent tools once limited to programmers.

Sources: langchain.com

Claude Code update makes Fable 5.2 the default Fable model

Version 2.1.321 of Claude Code, Anthropic's coding tool, sets Fable 5.2 as the default Fable model. It fixes a flaw where MCP sign-ins could send secrets to another host after a redirect. MCP is a standard for linking AI to outside tools. Other fixes cover permission prompts that failed to appear for some file writes and Bash commands.

Why it matters: Users should update to close security gaps as well as get the new model.

Sources: Anthropic on GitHub

LangChain offers LangSmith inside customers' own AWS accounts

LangChain made a bring-your-own-cloud version of LangSmith, its tool for monitoring and testing agents, generally available on AWS. Sensitive data such as traces and prompts stays in the customer's AWS account, while LangChain runs and upgrades the system. It is open to enterprise customers in 15 AWS regions across the US, EU and Asia-Pacific.

Why it matters: Regulated companies can track their agents without sending private data to an outside service.

Sources: langchain.com

SDKs + APIs

Amazon Bedrock knowledge base adds faster, simpler search modes

AWS says its Bedrock Managed Knowledge Base can now run single-shot retrieval-augmented generation, which searches documents once before an AI model writes an answer. A new retrieval-only mode returns matching passages without generating any answer. Both join the existing agentic mode that runs several searches, and developers pick among them with one setting.

Why it matters: Apps can answer easy questions quickly and cheaply while saving deeper searches for hard ones.

Sources: Amazon Web Services

Claude Fable 5.2 and Mythos 5.2 arrive on Amazon Bedrock

AWS says Fable 5.2 and Mythos 5.2 are now on Amazon Bedrock, with Mythos limited to Anthropic's verification programs. By default, using Fable 5.2 requires data retention for up to 30 days and human safety review by Amazon staff. Declined cybersecurity or biology requests can be retried on older Opus models, but developers' own code must handle that.

Why it matters: AWS customers need to plan for data retention rules before switching to the new model.

Sources: Amazon Web Services

GitHub Copilot adds Claude Fable 5.2 with opt-in for businesses

GitHub made Claude Fable 5.2 generally available in Copilot for Pro+, Max, Business and Enterprise users. Business and Enterprise admins must turn it on, because Anthropic keeps prompts and outputs to run safety checks. GitHub says eligible enterprises can use it with zero data retention through the end of the year.

Why it matters: Company admins have to weigh data retention terms before their developers can use the model.

Sources: The GitHub Blog

Nvidia releases setup recipes for Korean open AI models

Nvidia released deployment recipes for four Korean open models, including SK Telecom's A.X K2 and LG AI Research's K-EXAONE 2.0. They use Dynamo, Nvidia's open-source software for running models split across many GPUs, the chips most AI runs on. Nvidia says the recipes, tuned for its B200 chips, deliver up to twice the performance without extra tuning.

Why it matters: Running Korea's open models in production should now take less setup work.

Sources: NVIDIA Blog

Research

HOPE method trims AI models by keeping cooperating experts together

Researchers described HOPE, a way to remove unneeded experts from mixture-of-experts models, which send each input to a few specialized subnetworks. Earlier methods score each expert alone, but HOPE also considers pairs of experts that tend to work together. The authors report up to 6.1% better results on agentic coding tasks than the next-best method when half the experts are removed.

Why it matters: Smaller memory needs could let large models run on less hardware.

Sources: Hugging Face

Other notable news

Experts say Nvidia lacks a clear successor to Huang

Corporate governance experts told Axios that Nvidia has no obvious replacement if Huang has to leave. Axios notes Huang has dozens of direct reports, so no other executive oversees a broad part of the company. Founder ETFs, an Nvidia investor, told Axios it would sell its shares if Huang exits.

Why it matters: Relying so heavily on one leader leaves Nvidia's investors and customers exposed to sudden change.

Sources: Axios, The Information

Huang affair complicates Nvidia's push to sell chips to China

TechCrunch argues the affair weakens Huang's long-standing case that the US should keep selling AI chips to China. Nvidia earned $7.9 billion from China-based customers in the three months ended July 26, about 8% of revenue. A CSET analyst told The Verge a spy might seek insight into US export control thinking and who buys Nvidia chips.

Why it matters: Washington may now trust Nvidia's assurances about China sales less.

Sources: TechCrunch, The Verge

OpenAI builds Control Room to quickly halt research model jobs

OpenAI described Control Room, an internal system that tracks model jobs across its research infrastructure and can stop them. OpenAI says it halts output within 35 seconds of a decision to intervene, and stops output if Control Room itself fails. It works alongside monitors, separate models that watch training runs and raise alerts about problems.

Why it matters: AI labs are building emergency stops as their models take on more independent work.

Sources: OpenAI

Samsung, Kakao Mobility and LG expand physical AI work with Nvidia

Nvidia says Samsung cut one chip thermal simulation from about 18 hours to 41 seconds using Nvidia AI and GPU tools. Kakao Mobility plans to develop a level 4 autonomous driving system on Nvidia's vehicle computing platform. LG Electronics and Nvidia showed LG's CLOiD robot carrying out manufacturing tasks in a real factory.

Why it matters: Korean manufacturers are taking AI out of the lab and onto factory floors and roads.

Sources: NVIDIA Blog

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

---

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.