Data Points

Gemini 2.5 Pro June update now available in preview: Windsurf users seek alternatives to Claude 4 models

Mistral’s new integrated, fine-tunable coding tool. ChatGPT’s connectors, both official and DIY. Nvidia’s new small, open OCR and document analysis model. Reddit’s lawsuit against Anthropic over alleged scraping.

Analytics DeepLearning.AI

— 4 min read
Office team analyzing data and reports, with computers displaying graphs and charts in a modern workspace setting.

In today’s edition, you’ll learn more about:

  • Mistral’s new integrated, fine-tunable coding tool
  • ChatGPT’s connectors, both official and DIY
  • Nvidia’s new small, open OCR and document analysis model
  • Reddit’s lawsuit against Anthropic over alleged scraping

But first:

Google previews upgraded Gemini 2.5 Pro, spotlighting coding gains

Google introduced an enhanced preview of Gemini 2.5 Pro, which achieved a 24-point Elo score improvement on LMArena (reaching 1470) and a 35-point jump on WebDevArena (1443), while maintaining top positions on both leaderboards. The model tops coding benchmarks like Aider Polyglot, and demonstrates strong performance on GPQA and Humanity’s Last Exam, which test advanced math, science, and reasoning capabilities. Google also added “thinking budgets,” giving developers control over cost and latency trade-offs. The upgraded 2.5 Pro is available through the Gemini API at $1.25/$10 per million input/output tokens, and is rolling out in the Gemini app. (Google)

Anthropic cuts Windsurf’s access amid OpenAI acquisition rumors

Anthropic reduced AI coding startup Windsurf’s direct access to Claude 3.5 Sonnet and Claude 3.7 Sonnet models. The vibe coding company quickly found alternative third-party compute providers for the 3.x models, but still lacks access to Anthropic’s new Claude 4 models. Windsurf CEO Varun Mohan said the company wanted to pay for full capacity but was denied; Anthropic stated it was “prioritizing capacity for sustainable partnerships.” Anthropic co-founder Jared Kaplan later confirmed the decision was influenced by reports of OpenAI’s planned acquisition of Windsurf, stating it would be “odd for us to sell Claude to OpenAI,” Anthropic’s largest competitor. (TechCrunch)

Mistral launches enterprise coding assistant using open models

Mistral released Mistral Code, a coding assistant designed for enterprise software teams with heightened security needs. The platform combines four specialized models (Codestral for code completion, Codestral Embed for search, Devstral for complex coding tasks, and Mistral Medium for chat) into a single offering that can deploy in the cloud, on reserved capacity, or using air-gapped on-premises hardware. Mistral Code allows enterprises to fine-tune models on private repositories and keeps all code within the customer’s security boundary. The service entered private beta this week for JetBrains IDEs and VSCode, with general availability planned soon. (Mistral)

ChatGPT launches Connectors to integrate third-party apps and data sources

OpenAI introduced Connectors for ChatGPT, a beta feature that enables users to connect third-party applications like Google Drive, GitHub, and SharePoint directly into their conversations. The feature offers three types of connectors: chat search for quick file lookups, deep research for complex analysis across multiple sources, and synced connectors that pre-index content for faster responses. This integration allows AI developers to build more personalized workflows by accessing their own data sources without leaving ChatGPT, making it easier for individuals and teams to interact with their personal codebases and organizational knowledge. The feature is currently in beta, with Team, Enterprise, and Edu users having access to the widest range of services. (OpenAI)

Nvidia releases Llama Nemotron Nano VL for OCR and advanced document processing

Nvidia launched Llama Nemotron Nano VL, an open-weights vision-language model designed to extract information from documents like PDFs, charts, tables, and diagrams, while running on a single GPU. The model excels at document understanding tasks including question answering, table extraction, and visual element interpretation, achieving top performance on the OCRBench v2 benchmark for optical character recognition and document analysis. The model is available through Nvidia’s NIM API preview for download from Hugging Face. (Nvidia)

Reddit sues Anthropic over alleged unauthorized AI training

Reddit filed a lawsuit against Anthropic on Wednesday, accusing the AI company of training its models on Reddit users’ personal data without consent or compensation. Reddit alleges that Anthropic’s ClaudeBot scraped content in violation of Reddit’s user agreement, enabling the company to profit from its AI models. Anthropic’s competitors OpenAI and Google have both paid Reddit to license its users’ content. Reddit seeks a court injunction to stop the scraping, along with compensatory and punitive damages. (Ars Technica)

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 how non-engineers at AI Fund are learning to code with AI — starting with the ‘AI Python for Beginners’ course — and how this is empowering the entire team to build useful applications, boost creativity, and increase productivity.

“It is very empowering when individuals don’t have to try to get scarce engineering resources allocated to their ideas in order to try them out. There are a lot fewer gatekeepers in the way: If someone has an idea, they can build a prototype and try it out.”

Read Andrew’s full letter here.

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

Subscribe to Data Points

Read more