In today’s edition of Data Points, you’ll learn about our top headlines, and more:
- Claude Tag, your Slack coworker
- Seedance 2.5 improves video scene generation
- Robin, a biology-lab-in-the-loop
- Getty Images teams up with OpenAI
But first:
New custom chips designed for OpenAI LLMs
OpenAI and Broadcom announced Jalapeño, OpenAI’s first custom chip designed specifically for large language model inference. The chip went from initial design to the manufacturing foundry (what the industry calls “tape-out”) in nine months, partly by using OpenAI’s own models to accelerate design and optimization. Early internal testing shows Jalapeño delivers better performance per watt than current accelerators by reducing data movement and balancing compute, memory, and networking around the specific patterns of inference workloads. Engineering samples are already in production in the lab, including GPT-5.3-Codex-Spark. Deployment begins with Microsoft and other data center partners by the end of 2026, with future generations planned as part of OpenAI’s push to control more of its infrastructure and drive down costs. (OpenAI)
Fugu approaches Claude Fable 5 performance by blending models
Sakana AI released Fugu, a multi-agent system that routes queries across specialized models and synthesizes results behind a single API. The timing is telling: CEO David Ha framed it explicitly as insurance against vendor lock-in and geopolitical risk. Fugu breaks complex requests into sub-tasks, delegates them to different foundation models, and stitches the answers together. The orchestration happens behind the scenes; Sakana keeps its routing logic proprietary but lets enterprises block specific providers for compliance or to prevent prompts from feeding training data. On SWE-Bench Pro, Fugu Ultra solved 73.7 percent of problems, beating Claude Opus 4.8 (69.2) and GPT-5.5 (58.6)., though the now-inaccessible Fable 5 still tops it at 80.0. At $5 per million input tokens, $30 per million output, Fugu’s price is well above mid-tier models. The service isn’t available in the EU yet while the company figures out how to square its black-box architecture with GDPR. (VentureBeat)
Claude upgrades its Slack integration for AI teamwork
Anthropic released Claude Tag, a Slack integration that lets teams tag “@Claude” as a teammate to handle tasks across channels. Unlike single-chat interfaces, Claude Tag maintains shared context within a workspace: Anyone can see what Claude is working on, hand off tasks mid-conversation, and Claude learns from channel history to avoid redundant explanations. The model can also work asynchronously, scheduling its own follow-ups over hours or days, or be granted access to specific tools and data sources with tight administrative controls. Anthropic says its internal version already powers 65 percent of product team code generation. Claude Tag launched Tuesday in beta for Enterprise and Team customers on Slack, replacing the existing Claude in Slack app. (Anthropic)
Bytedance video model can synthesize more references
Seedance unveiled version 2.5 of its AI video generator, an upgrade focused on two practical improvements: higher-resolution output and extended video duration. The update targets users who need sharper detail for product visuals, social ads, and portfolio work who don’t want to overhaul their existing workflow. Longer clips now support complete scenes—a reveal, camera move, or character beat—rather than cutting off mid-moment. The generator maintains Seedance’s familiar prompt-and-refine process but adds aspect ratio and duration controls tuned for specific platforms. For users working toward publishable video, the resolution bump and scene continuity improvements mean less upsampling and fewer stitched clips to manage. (Seedance)
Autonomous agents come to biological research
Researchers introduced Robin, a multi-agent AI system that fully automates hypothesis generation, experimental design, data analysis, and data interpretation in experimental biology. The system integrates literature search agents with data analysis agents to propose experiments, interpret results, and refine hypotheses in a closed loop. Robin identified one therapeutic candidate for dry age-related macular degeneration (a condition that causes partial or total blindness): Ripasudil, a clinically approved glaucoma drug. A follow-up analysis designed by Robin also revealed a novel therapeutic target (ABCA1). All hypotheses, experimental directions, analyses, and figures were generated by the system itself, marking the first autonomous discovery and validation of novel drug candidates within an iterative “lab-in-the-loop” framework. (Nature)
OpenAI partners with Getty for well-attributed images
Getty Images announced a display partnership with OpenAI that will display the company’s licensed images across ChatGPT’s search and discovery features. The agreement represents Getty’s effort to position itself as a trusted, fully-licensed content source for AI applications, arguably a notable shift as AI companies face ongoing legal challenges over training on unlicensed image data. Getty CEO Craig Peters suggested that licensed imagery makes AI-powered search more reliable than alternatives. The parties didn’t reveal financial terms. (GlobeNewswire)
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 U.S. Government and Anthropic’s actions to control AI access, Anthropic’s release of Claude Fable 5 with restrictive guardrails, and the implications for AI sovereignty and open-source alternatives.
“Platforms succeed when they are viewed as stable, reliable partners that one can build on. The sudden rule changes by Anthropic (including a mandatory 30-day data retention policy for Fable usage) have made developers wonder about the stability of building on any one proprietary LLM provider, not just Anthropic.”
Read Andrew’s letter here.
Other top AI news and research stories covered in depth:
- Independent tests of Claude Fable 5 have encountered Anthropic’s protective policies, raising questions about transparency in AI benchmarking.
- New agentic tests like DeepSWE, ProgramBench, and ITBench-AA are pushing AI agents beyond traditional bug hunts, setting new standards for evaluation.
- Nvidia is making a bold move with Nemotron 3 Ultra, focusing on speed and openness to capture the competitive AI hardware market.
- Researchers have developed Privileged On-Policy Exploration (POPE), a reinforcement learning technique that trains models to expand on partial solutions, enhancing problem-solving capabilities.
Our 7-Day Voice AI Builder Challenge
Stop watching the terminal. Teach your agent to call you when it matters.
DeepLearning.AI and Vocal Bridge are teaming up to host a 7-day Voice AI Builder Challenge. Your mission will be to teach your AI coding assistant to call you the moment it needs human intervention.
Compete for:
- Top-tier prizes for the highest-scoring builds
- Real-time expert feedback to help you refine your assistant on the fly
- A spot on our live, competitive leaderboard
- Global visibility, with the winning skill featured on the official DeepLearning.AI GitHub repo
The waitlist is officially open. Join today to ensure you're notified the exact moment the challenge goes live.
Data Points is produced by human editors with AI assistance.