Claude Fable and Mythos trim the verbiage: Perplexity preserves privacy on local machines

In today’s edition of Data Points, you’ll learn about our top headlines, and more:

• World Labs founder Fei-Fei Li introduces a world model
• Times-FM-3 sets new standard for multivariable time forecasts
• ChatGPT’s advertising program expands to new countries
• Music publishers sue Anthropic for copyright infringement

But first:

Claude’s flagship models get a solid upgrade

Anthropic made Claude Fable 5.1 generally available alongside Claude Mythos 5.1, a version with fewer safeguards available only through trusted access programs for cybersecurity and life-sciences work. Anthropic says Fable 5.1 costs about 25 percent less than Fable 5 for typical workloads and up to 45 percent less for agentic tasks, due to cheaper cache-read pricing. Fable 5.1 beats Fable 5, Opus 5, and GPT-5.6 Sol on benchmarks including Terminal-Bench-Science (52.6 percent versus 24.7 percent for Fable 5) and CursorBench (73.4 percent versus 70.5 percent). Cybersecurity safeguards also now produce 60 percent fewer false positives, and the model’s writing is less verbose. Anthropic also introduced Enterprise Frontier Safeguards, a program that stores customer data on infrastructure the customer controls instead of Anthropic’s cloud. (Anthropic)

Perplexity agent blends local and cloud models to preserve privacy

Perplexity launched Hybrid Compute for its Perplexity Computer agent on Mac. It starts tasks using cloud-based frontier models and hands off any step touching private files to a compact model running locally, then merges the results. The feature works on Apple Silicon Macs with macOS 15 or later and at least 24GB of unified memory (32GB recommended; it’s available to Pro, Max, and Enterprise subscribers. It installs with one click and uses no cloud credits for local work. An on-device classifier called PII-Tracer, a 0.6-billion-parameter model Perplexity open-sourced along with a benchmark named PII-TRACE, decides whether sensitive data stays local, gets masked, gets refused, or requires user consent, handling credentials, payment numbers, and government IDs most strictly. Perplexity reports PII-Tracer leads 12 detectors on character-level F1 (0.629) and on catching every instance of a recurring identifier (79.4% versus 57.0% for the next-best model, GPT-5.6-sol), but recall drops sharply on long conversations before a sliding-window decoding fix restores it. For developers, Hybrid Compute offers a pattern for building agents that handle sensitive local files without routing everything through a cloud API, backed by an open-source PII detector others can inspect or reuse. (MarkTechPost)

World Labs’ Atlas lays a foundation for world-building

Fei-Fei Li’s World Labs released Atlas, a new model trained to handle text, images, video, and 3D data in one architecture. It generates images and video with camera control, producing up to one minute of 1440p video, and reconstructs real scenes into 3D outputs like point clouds and Gaussian splats from as few as two or three input photos. Atlas is a multimodal autoregressive diffusion transformer (a model that generates step by step while also using diffusion techniques common in image generators) and borrows serving techniques from large language models and video diffusion models. The company says Atlas outperforms specialized models on camera-conditioned generation and 3D reconstruction benchmarks, but did not publish its full comparison methodology. Atlas will power future versions of World Labs’ Marble product, aimed at VFX, robotics simulation, and game design work that currently needs separate reconstruction and generation tools. (World Labs)

Google updates its forecasting model, for noncommercial use only

Google Research released TimesFM-3, a 330-million-parameter foundation model that forecasts multiple related time series in a single forward pass without task-specific fine-tuning. The model can be used to predict events in fields including retail, finance, observability, manufacturing, healthcare, and natural sciences. Earlier TimesFM versions only handled one series at a time using its own history; TimesFM-3 is pretrained natively on more than 1 trillion time points and accepts multiple targets, past covariates, and past-future covariates like promotion calendars. It uses a decoder-only transformer with alternating causal temporal attention and full variate attention, plus a masking technique called Contiguous Patch Masking that produces the whole forecast horizon in one pass with nine quantiles per step. Google reports it ranks first among pretrained foundation models on the GIFT-Eval, fev-bench, and TIME benchmarks, beating Chronos-2 and Toto 2.0. The model weights carry a non-commercial license restricting production use, so developers who need to ship a forecasting product must still use the Apache-2.0-licensed TimesFM 2.5 instead. (MarkTechPost)

Ad-supported chat pays off for OpenAI

OpenAI’s ChatGPT Ads reached $1 billion in annualized revenue run rate less than 200 days after launch, with tens of thousands of advertisers now using the platform. Starting the day of the announcement, self-service ad purchasing through Ads Manager expands to India, Europe, the Middle East, and North Africa, adding to availability in more than 40 countries. Ads are shown based on the context of a user’s current conversation and, depending on settings and country, their broader ChatGPT usage. OpenAI says ads are labeled, don’t influence ChatGPT’s answers, and advertisers can’t access private conversations. The platform now supports cost-per-click and outcome-based bidding, product feeds, Pixel and Conversions API tracking, and custom audience targeting, with SMBs making up a growing share of advertisers since Ads Manager launched in May. OpenAI reports one advertiser saw 3x return on ad spend over 28 days, though it did not disclose broader performance data across the platform. OpenAI credits ads with helping expand access to its models worldwide. (OpenAI)

Anthropic sued (again) for training on song lyrics

Sony Music Publishing and Warner Chappell Music sued Anthropic, along with CEO and co-founder Dario Amodei and co-founder Benjamin Mann, in the U.S. District Court for the Northern District of California, alleging Claude was trained on tens of thousands of copyrighted songs without permission. The complaint accuses Anthropic of torrenting pirated books from Library Genesis and Pirate Library Mirror, scraping lyrics from licensed sites like Musixmatch and LyricFind, and running a “destructive scanning” operation on physical books, citing findings already unsealed in the separate Bartz v. Anthropic authors’ case. The publishers seek statutory damages up to $150,000 per willfully infringed work plus up to $25,000 per removed copyright notice, destruction of infringing copies, and disclosure of Claude’s training data. This is the fifth major music-industry suit against Anthropic, following actions from Universal Music Publishing Group, Concord, and ABKCO (over $3 billion), BMG, and Round Hill Music. It also follows Anthropic’s $1.5 billion settlement with book authors over similar torrenting conduct, announced in 2025 and approved this year. The suit signals that music publishers are pushing for damages and licensing terms rather than accepting a one-time settlement. (Music Business Worldwide)

Want to know more about what matters in AI right now?

Read the latest issue of The Batch for in-depth analysis of news and research.

Last week, Andrew Ng talked about the continued importance of understanding software engineering fundamentals in the context of agentic coding, emphasizing the ability to build full-stack applications, manage data, and design system architectures to make informed tradeoffs and ensure secure, reliable, and scalable systems.

“Even when you use a coding agent to write all your code, understanding software fundamentals is important for steering your agent to make the tradeoffs you want — or to even know what tradeoffs exist to be made. Additionally, when you’re building an AI application, the AI core is often expressed through a broader software application, which a skilled engineer will shape.”

Read Andrew’s letter here.

Other top AI news and research covered in depth:

• Z.ai delayed its open weights release for GLM-5.3 due to cybersecurity risks, marking a new approach to AI safety protocols.
• DeepSeek’s flagship model, DeepSeek-V4-Pro, only slightly outperforms its Flash sibling, but its open source agent harness is drawing significant interest from the AI community.
• OpenAI partners with Cerebras, Google releases Gemini 3.7 Flash, and Nvidia debuts Nemotron 3.5 Lightning, highlighting the AI industry's ongoing race for speed.
• Researchers at Xiaohongshu have developed an LLM-based memory management technique for agents, potentially transforming how AI systems handle data retention and processing.

A special offer for our community

DeepLearning.AI’s first-ever subscription plan for our entire course catalog includes foundational classics like the Machine Learning and Deep Learning specializations plus seminars on the latest tools and frameworks you need.

As a Pro Member, you’ll immediately enjoy access to:

• Nearly 200 short and long AI courses from Andrew Ng and industry experts
• Labs and quizzes to test your knowledge
• Projects to share with employers
• Certificates to testify to your new skills
• A community to help you advance at the speed of AI

Enroll now to lock in a year of full access for $25 per month paid upfront, or opt for month-to-month payments at just $30 per month. Both payment options begin with a one-week free trial.

Explore Pro’s benefits and start building today!

Try Pro Now!