In today’s edition of Data Points, you’ll learn about our top headlines, and more:
- Claude Science, a workbench for biological research
- Gemini’s Nano Banana 2 Lite and Omni Flash
- Devin Fusion pairs a top model with an inexpensive sidekick
- OpenClaw’s iOS and Android controller applications
But first:
Anthropic restores Claude Fable 5 access
The Trump administration is lifting export restrictions on Anthropic’s Claude Fable 5 and Claude Mythos 5 after the company reached a deal with the Commerce Department. The move resolves a months-long dispute over jailbreak vulnerabilities that could let users bypass safety restrictions, particularly around cybersecurity capabilities. Anthropic initially argued that eliminating all jailbreaks was technically impossible, but shifted its approach in recent weeks, replacing CEO Dario Amodei in negotiations with cofounder Tom Brown and committing to build stronger safeguards instead. Commerce Secretary Howard Lutnick communicated the decision to Anthropic in a letter on Tuesday and Anthropic began restoring access to Claude Fable 5 on Wednesday. The resolution opens the way for broader access to Mythos, which had been restricted to select companies and government agencies, and restores Fable after it was pulled offline. (Wired)
Claude Sonnet 5 arrives at a temporary discount
Anthropic launched Claude Sonnet 5, a mid-tier model designed for autonomous agent work that performs close to the more expensive Claude Opus 4.8 while costing less. The model shows substantial improvements over Claude Sonnet 4.6 in reasoning, tool use, coding, and multi-step task completion. Early testers reported Claude Sonnet 5 finishing complex tasks where earlier Sonnet versions would stall, checking its own work unprompted, and handling everything from software engineering to legal research to insurance workflows. Pricing starts at introductory rates of $2 per million input tokens and $10 per million output tokens through August 31, 2026, then moves to $3 and $15 respectively.. The model launches with cyber safeguards enabled by default and shows lower rates of undesirable behaviors than its predecessor, though it remains somewhat more capable of performing misaligned tasks than Claude Opus 4.8. (Anthropic)
Anthropic pulls together its life science tools into one platform
Anthropic released Claude Science, a unified AI research workbench that brings together tools like PubMed, Jupyter, and connections to HPC clusters into a single environment for scientific work. into a single app for scientific work. The platform ships with over 60 pre-configured skills spanning genomics, proteomics, and structural biology, plus a reviewer agent that checks citations and calculations. It runs locally or on remote infrastructure, so sensitive datasets never leave a lab’s existing systems, and generates fully reproducible artifacts with auditable histories showing exactly how each figure or analysis was created. Early users include Manifold Bio, which used it to rank drug targets against proprietary criteria, and Allen Institute neuroscientist Jérôme Lecoq, who built a multi-agent pipeline that compressed a two-year literature review process into something fast enough to produce ten reviews of over 100 pages each. (Anthropic)
Gemini updates image and video generation models
Google released two new generative media models for developers: Nano Banana 2 Lite, a lightweight image generation model optimized for speed and cost, and Gemini Omni Flash, a video generation and editing model that works with multimodal inputs. Nano Banana 2 Lite generates images in 4 seconds at $0.034 per 1K-resolution image, positioning it as a replacement for slower, more expensive alternatives. Gemini Omni Flash handles video generation and conversational editing at $0.10 per second of video output, supporting 10-second clips with longer durations coming soon. Both models are available immediately across Google AI Studio, the Gemini API, and consumer products like Search and the Gemini app. Google argues that the real value lies in chaining them together: use Nano Banana 2 Lite to generate images at scale, then feed those images to Omni Flash to animate them into video. Google published three demo apps showing the workflow, including one that lets users photograph themselves and instantly transport them to iconic landmarks as animated clips. (Google)
Devin Fusion uses model orchestration to reduce costs
Cognition released Devin Fusion, a multi-model system that routes tasks between a high-performing AI model and a cheaper “sidekick” model while maintaining frontier-level performance on code generation tasks. The architecture runs two parallel agents—one with an expensive model handling planning and final review, the other with a cost-effective model handling routine work—and both maintain their own cached contexts to avoid expensive cache misses when switching models. On FrontierCode, Cognition’s coding benchmark measuring real-world quality, Fusion achieves 35 percent lower costs than leading models like GPT-5.5 and Opus 4.8 while matching their performance. The system adds dynamic mid-session routing: lightweight classifiers monitor task difficulty during execution and can switch models or upgrade the sidekick without extra cache penalties by timing changes during context compaction. Internal testing showed 88 percent of Cognition’s merged pull requests were handled entirely by the automated router, suggesting the approach works on actual engineering tasks rather than just benchmarks. (Cognition)
OpenClaw agents come to smartphones
OpenClaw released free companion apps for iOS and Android that connect phones to a self-hosted AI agent Gateway running on a separate machine. The phones act as peripheral nodes, not standalone assistants, exposing device hardware like camera, location, and voice to the Gateway over WebSocket. The separation is intentional: the Gateway controls all sessions and routing, while nodes simply add sensor access and a Canvas interface for dashboards. Pairing happens via QR code or setup code, and sensitive commands like camera access default to off until explicitly allowlisted in the config. The architecture works on local networks via mDNS or remotely through Tailscale with encrypted WSS endpoints, giving developers a way to build workflows that can see, hear, and navigate the physical world while keeping the core agent logic centralized and secure. (MarkTechPost)
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 talked about the concept of “loop engineering” and its application in AI-driven software development, highlighting three key loops—agentic coding, engineering, and developer feedback loops—that enhanced coding efficiency and product management.
“AI-native teams are increasingly using AI to help shape product direction, for example, automating the gathering and analysis of usage data, summarizing written and verbal customer feedback, or carrying out competitive analysis. However, for pretty much all the products I’m involved in, I see humans as having a significant context advantage over current AI systems — we know a lot more than the AI system about the users and the context the product has to operate in — and thus humans play a critical role.”
Read Andrew’s letter here.
Other top AI news and research stories covered in depth:
- Top Agentic Performance, Low Cost highlights GLM-5.2’s efficiency, delivering high performance rivaling that of closed models.
- AI Degrees on the Rise explores the increasing availability of AI-focused programs in U.S. universities, from comprehensive majors to specialized minors.
- Large-Model AI for Apple Devices details Apple’s 2026 initiative to integrate advanced AI models into MacBooks, iPhones, and cloud services.
- Biological Molecules as Language introduces ESMFold2, a new Transformer-based architecture that rivals AlphaFold 3 in predicting biological structures.
A special offer for our community
In case you missed it, DeepLearning.AI launched our first-ever subscription plan for our entire course catalog! As a Pro Member, you’ll immediately enjoy access to:
- Nearly 200 AI short and long 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!
Data Points is produced by human editors with AI assistance.