AI News Summary - Week 19.2, 2026

Another busy week in AI brings major enterprise moves, groundbreaking voice capabilities, and historic achievements in space AI.

OpenAI Unveils GPT-Realtime-2: Voice Intelligence Enters a New Era

OpenAI has released three new audio models that represent a significant leap in voice AI capabilities:

• GPT-Realtime-2: The first voice model with GPT-5 class reasoning, capable of handling complex requests while maintaining natural conversation flow.
• GPT-Realtime-Translate: Live translation between 70+ input languages and 13 output languages.
• GPT-Realtime-Whisper: Real-time speech-to-text that transcribes as users speak.

The models extend context windows to 128K and introduce adjustable reasoning effort. Early testing shows a 26-point improvement in call success rates at Zillow, demonstrating the model's suitability for production voice agents.

Enterprise AI Wars: Anthropic and OpenAI Launch Major Joint Ventures

Both Anthropic and OpenAI announced parallel enterprise-focused joint ventures this week:

Anthropic's $1.5B Venture

• Backed by Blackstone, Hellman & Friedman, Goldman Sachs, Apollo, General Atlantic, GIC, and Sequoia Capital
• Each major partner committing $300 million
• Focus on deploying AI with forward-deployed engineers in mid-sized companies

OpenAI's $10B "Development Company"

• Raising $4 billion from 19 investors including TPG, Brookfield, Advent, and Bain Capital
• Targeting enterprise AI deployment at scale
• Valuation reported at $10 billion

Both companies are racing toward potential IPOs, with Anthropic reportedly seeking a $900 billion valuation and OpenAI recently raising$122 billion at $852 billion valuation.

NASA Makes History: First AI Foundation Model Deployed in Orbit

NASA and IBM's open-source Prithvi Geospatial AI model has become the first geospatial foundation model deployed in orbit. Researchers from Adelaide University successfully uploaded a compressed version to:

• The South Australian government's Kanyini satellite
• The IMAGIN-e payload aboard the International Space Station

The model, trained on 13 years of satellite data from Landsat and Sentinel-2, can now perform flood and cloud detection directly in space. This demonstrates how foundation models could revolutionize Earth observation by enabling multiple tasks with minimal bandwidth updates.

DeepMind's AlphaEvolve Achieves Cross-Domain Breakthroughs

Google DeepMind's AlphaEvolve coding agent, powered by Gemini, is delivering remarkable results across multiple fields:

• Genomics: 30% reduction in DNA sequencing errors at PacBio
• Grid Optimization: Increased feasible solution rates from 14% to 88% for power flow problems
• Disaster Prediction: 5% accuracy improvement across 20 natural disaster categories
• Quantum Computing: 10x lower error rates for quantum circuits on Google's Willow processor
• Mathematics: Solved Erdős problems with Terence Tao and broke records on Traveling Salesman and Ramsey Numbers

Perhaps most notably, AlphaEvolve has been integrated directly into the silicon of Google's next-generation TPUs with "counterintuitive yet efficient" circuit designs.

Sources: TechCrunch, OpenAI, NASA, Google DeepMind