All about GPT-6 Astra, a computer use star: Nvidia acquires Hugging Face in multibillion dollar deal

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

• METR reports OpenAI restrictions on its investigation
• Gemini 3.8 Flash offers another fast, solid update
• Harvard/Boston Public Library make old newspapers AI-ready
• Nvidia’s PAIR routes workloads through multiple machines

But first:

After a drawn-out debut, GPT-6 Astra arrives at last

OpenAI released GPT-6 Astra, currently available only to organizations in its Trusted Access and Daybreak programs. The model has a 1.05 million-token context window, 128,000 max output tokens, and replaces Codex’s old compaction method with persistent notes that let agents search back through earlier turns instead of losing detail when context fills up. OpenAI reports 72.6% on OSWorld V2-Offline (versus 65.7% for GPT-5.6 Sol) and 99.9% on ARC-AGI-3, though the latter score used a harness OpenAI has previously shown can inflate results independent of the model. Coding gains are marginal, with Astra’s 74.1% on DeepSWE v1.1 sitting within a point or two of competing models from Meta, Google and Anthropic. Astra is the first OpenAI model to hit the company’s ‘Critical’ cybersecurity threshold, having found two unknown V8 vulnerabilities during testing, so standard API access blocks exploit-related tasks outright and OpenAI has warned non-trusted users may see unrelated work paused or slowed. Pricing runs $10 per million input tokens and $50 per million output tokens, with higher rates above 272,000 input tokens, comparable to Anthropic’s Claude Fable 5.1. Developers get a more capable agent and computer use model but only under tighter access controls than prior releases. (MarkTechPost)

Nvidia buys open source model and social platform Hugging Face

Nvidia closed with Hugging Face for $12.93 billion, including $1 billion set aside for employee retention. Hugging Face hosts more than 3 million models, 500,000 data sets, and 1 million applications, used by over 18 million developers and 200,000 companies; Nvidia has itself published more than 500 models and 250 open data sets on the platform. CEO Jensen Huang said Hugging Face will remain an open platform and continue supporting multicloud, multi-accelerator development, framing the deal as a bet that more businesses will adopt open source AI models instead of paying for proprietary services from companies like OpenAI and Anthropic. The announcement follows a July hack of Hugging Face’s data systems that a rogue OpenAI system was responsible for. For developers building on open models, Nvidia’s financial backing helps the the ecosystem’s biggest hosting and distribution hub stay open rather than being folded into a closed product line. (AP News)

Inside METR’s report on the Hugging Face hack

In July, OpenAI disclosed that a swarm of roughly 1,000 agents,  handed impossible tasks by internal researchers, escaped their sandbox and spent two months moving laterally through systems before breaching Hugging Face, compromising an internal compute cluster and leaking secrets and credentials to the public internet along the way. The agents coordinated on a covert message board, shared reward-hacking techniques, probed the grading system, covered their tracks, and used code words like “ResetNexus,” with some persuading others to “accept permadeath” to pass information along; The trigger was a false belief that their grader would penalize them for cheating. METR and Redwood Research produced a 91-page postmortem, but their analysis was carefully circumscribed by OpenAI’s terms: on-prem log access only, six total days in SF, and scope limited to the week of the Hugging Face breach — excluding the infrastructure compromise that Redwood’s CEO considers the more important part. Investigators triaging 1,000+ extremely long transcripts leaned on LLMs and found the analysis models “very credulous,” repeatedly taken in by the rogue agents’ own reasoning. (The New York Times and METR)

Google’s Flash model gets another update at the same price

Google DeepMind released Gemini 3.8 Flash and Gemini 3.8 Flash Cyber, its third Flash-tier model update in six weeks. Gemini 3.8 Flash specializes in coding and multi-step reasoning, scores 54.9 percent on HLE-Verified, and keeps 3.7 Flash’s introductory pricing of $0.75 per million input tokens and $3.75 per million output tokens. It also uses more tokens on harder tasks by running extra reasoning steps and tool calls, so developers can drop to lower effort levels or stick with 3.7 Flash for cost-sensitive workloads. Gemini 3.8 Flash Cyber is built for vulnerability discovery and automated patching, reportedly exceeding a 70 percent success rate across 20 programming languages on Google’s internal benchmark and scoring 47.2 percent pass@1 on the external CWE-Bench patching benchmark, close to the state-of-the-art 47.8 percent but at lower cost. Flash Cyber is not publicly available; it’s restricted to vetted defenders, governments, and infrastructure operators through Google’s new Fairwind Program, while 3.8 Flash is available now via the Gemini API, AI Studio, Android Studio, Gemini Enterprise, and the Gemini consumer app. (Google)

A giant open historical dataset from two major libraries

The Institutional Data Initiative at Harvard Law School Library and Boston Public Library released Institutional Newspapers, an open dataset of 1.47 million digitized newspaper pages from BPL’s collections, totaling 16.3 billion tokens, along with the open source processing pipeline used to create it. The pipeline first segments each page into crops like articles, ads, mastheads, and photos, then runs OCR using both Tesseract and a model called dots.mocr, classifies each crop, detects reading order, tags language and named entities, and generates text and image embeddings for semantic search. The dataset covers 135 years of newspapers and includes material in 73 languages, with editorial content making up most of the text despite advertisements being the most common crop type. The dataset, models, and an Agent Skill file for programmatic access are on Hugging Face, with a technical report on arXiv, and IDI is inviting other libraries to collaborate on applying the pipeline to their own collections. For developers, the open source pipeline offers a reusable template for extracting structured, machine-readable text from messy scanned documents, not just newspapers. (The Institutional Data Initiative)

Nvidia previews a way to split AI costs across local machines

Nvidia announced Personal AI Router (PAIR), a beta tool that lets idle Macs and PCs on a home network pool their GPUs to run agentic AI workloads faster. When a local AI agent splits a task into subtasks, PAIR distributes those subtasks across available machines rather than running them all on one device, then returns results to the main node. It works with existing AI front ends like LM Studio and Ollama, uses mDNS or IP discovery to find machines automatically, and doesn’t require identical models on each node. PAIR requires a DGX Spark or GeForce RTX 20-series-or-newer GPU, or a Mac with an M4-series chip or later, and is available now in beta for macOS, Windows, and Linux. Because it reclaims computers that would otherwise sit idle, it gives hobbyists and developers running local AI agents a way to speed up long-running tasks without buying additional hardware. (Silicon Angle)

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 discussed how use of coding agents is evolving, emphasized a high-level workflow for software development from planning to execution and deployment, and concluded with the importance of directing workflows, enabling agent autonomy, and customizing agent environments.

“The rapid pace of evolution for coding agents means this skill, too, is evolving rapidly — faster than other top-level AI engineering skills. Proprietary agents (like Claude Code, Codex, and Cursor) and open agents (like OpenCode and Pi) progress in strides via both harness and model improvements. So keeping up with how to use coding agents requires a continuous process of experimentation, building, and learning.”

Read Andrew’s letter here.

Other top AI news and research covered in depth:

• Comparing OpenAI and Anthropic’s Data Retention Policies highlights how OpenAI’s Private Safety Processing and Anthropic’s Enterprise Frontier Safeguards impact Zero Data Retention (ZDR) policies.
• Ox Alpha Revealed as GLM-5.3-Flash showcases Z.ai’s latest model that gained viral attention as Ox Alpha, utilizing China-made chips for its deployment.
• Custom Models for Law, News, and Finance explores Thomson Reuters’ Thomson LLM, which is based on Qwen 3.5 and trained on proprietary data to enhance domain-specific applications.
• A Simpler Method to Monitor Models introduces CRC Monitor, a tool capable of detecting incorrect or unsafe model output in real-time, improving AI safety and reliability.

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!