Google’s latest Gemma model is a power-saver: Claude boosts Sonnet 4’s input limits

Analytics DeepLearning.AI -

In today’s edition of Data Points, you’ll learn more about how:

  • Building a more interpretable robot
  • OpenAI’s latest gold medal programming performance
  • Extracting structured data from big texts
  • Anthropic adds learning modes to Claude Code

But first:

Google releases Gemma 3 270M, a tiny model optimized for mobile devices

Google’s 270 million parameter model (170 million embedding parameters and 100 million transformer parameters, with a relatively large 256,000-token vocabulary) is designed for task-specific fine-tuning, with built-in instruction-following and text structuring. Gemma 3 270M achieves strong performance on the IFEval benchmark for instruction-following despite its small size. Internal tests also show a 4-bit quantized version consumed just 0.75 percent of a Pixel 9 Pro’s battery across 25 conversations, making it Google’s most energy-efficient Gemma model. For developers who need lightweight, specialized AI models for high-volume tasks like sentiment analysis, entity extraction, and content moderation, this release offers an alternative to larger general-purpose models. The model is available for free through Hugging Face, Ollama, Kaggle, and other platforms, with both base and instruction-tuned versions included. (Google)

Claude Sonnet 4 expands context window to 1 million tokens

Anthropic increased Claude Sonnet 4’s context window from 200,000 to 1 million tokens, enabling developers to process entire codebases in a single API request. The expanded context window supports large-scale code analysis, document synthesis across hundreds of files, and context-aware agents that maintain coherence across extensive tool calls and workflows. Pricing doubles for prompts exceeding 200,000 tokens, with input costs rising from $3 to $6 per million tokens and output costs increasing by 50 percent, from $15 to $22.50 per million tokens. This boost lets applications built on Claude Sonnet 4 handle significantly more complex, data-intensive tasks – over 75,000 lines of code or dozens of research papers – while maintaining full context awareness. The feature is now in public beta for selected developers on the Anthropic API and in Amazon Bedrock, with Google Cloud’s Vertex AI support coming soon. (Anthropic)

MolmoAct introduces action reasoning models for more explainable robot control

Researchers from the University of Washington and Allen Institute for AI have developed MolmoAct, a family of open-source robotic foundation models that integrate perception, planning, and control through structured reasoning. The models generate three types of tokens sequentially: depth perception tokens for 3D understanding, visual reasoning traces showing planned trajectories, and action tokens for robot control. MolmoAct-7B-D achieved 70.5 percent zero-shot accuracy on SimplerEnv Visual Matching tasks, surpassing closed-source models π0 and GR00T N1 (while taking much less time to pre-train), and 86.6 percent average success on LIBERO benchmarks. This more transparent approach to model trajectories in particular addresses some limitations in current vision-language-action models, making robot decision-making more explainable and steerable through visual trajectory editing. The team released all model weights, training code, and the MolmoAct Dataset containing over 10,000 robot trajectories. (arXiv)

OpenAI’s AI system wins gold at IOI programming olympiad

OpenAI’s AI reasoning system scored at a gold-medal level at the 2025 International Olympiad in Informatics, ranking sixth among 330 human contestants and first among AI entrants. The system competed under identical constraints as human participants, including a five-hour time limit, submission caps, and no internet access, using only a basic terminal environment. OpenAI’s ensemble of general-purpose models, which weren’t specifically trained for the competition, improved from the 49th percentile in 2024 to the 98th percentile in 2025. The models’ performance shows significant progress in AI’s ability to solve complex programming problems under standardized conditions, potentially informing future development of coding assistants and developer tools. (X/Twitter)

Google releases LangExtract, an open-source Python library for extracting structured data from unstructured text

LangExtract uses large language models to extract structured information from unstructured documents, including anything from clinical notes and radiology reports to classic literature. The library maps every extraction to its exact source location in the text, generates interactive HTML visualizations for reviewing results, and supports both cloud-based models like Gemini and local models through Ollama. LangExtract handle long documents effectively, using an optimized approach with text chunking, parallel processing, and multiple extraction passes while requiring only a few examples to define custom extraction tasks for any domain. LangExtract is one of several tools addressing a critical need in AI development for reliable, traceable information extraction from large bodies of text without requiring model fine-tuning. The library is not officially supported by Google, but is available on PyPI and GitHub under an Apache 2.0 license. (GitHub)

Claude expands learning modes to guide users through problems

Anthropic added new learning modes for its Claude AI assistant that emphasize guided discovery and step-by-step reasoning rather than providing immediate solutions. The features were originally developed for the education market but are now available for both Claude.ai and Claude Code. “Explanatory” and “learning” modes in Claude Code use Socratic questioning and collaborative problem-solving approaches, pausing mid-task to ask developers to initiate code sections. The learning modes work through modified system prompts rather than fine-tuned models, allowing rapid iteration based on user feedback while addressing the challenge of maintaining educational value alongside productivity gains. The launch follows similar educational features from OpenAI (Study Mode) and Google (Guided Learning), reflecting industry-wide concerns that students and junior developers could become overly dependent on AI-generated answers without understanding underlying concepts. (VentureBeat)

Still want to know more about what matters in AI right now?

Read this week’s issue of The Batch for in-depth analysis of news and research.

This week, Andrew Ng shared his experience visiting the University of Exeter in the UK to receive an honorary doctorate, highlighting the university leadership’s enthusiastic embrace of AI and its forward-looking approach to integrating AI across disciplines like computer science, environmental science, and business.

“This is not a group whose primary worry is whether students will cheat using AI. This is a group that is thinking about how to create a student body that is empowered through AI, whether by teaching more students to code, helping them use AI tools effectively, or showing them what’s newly possible in their disciplines.”

Read Andrew’s full letter here.

Other top AI news and research stories we covered in depth:

Subscribe to Data Points