Daily AI RoundupsFriday, 31 July 2026

The 31 July 2026 Calathea roundup leads with OpenAI cutting GPT-5.6 Luna prices by 80% and Terra by 20%, alongside a Sol API mode offering up to 2.5 times the speed at twice the standard price. The EU opens tenders for up to seven AI gigafactories, while Google DeepMind launches Gemini Robotics ER 2 and DeepSeek puts V4 Flash into public beta with an 82.7 Terminal-Bench 2.1 score. MiniMax introduces H3, Thinking Machines releases the 276B-parameter Inkling-Small model and Huawei publishes openPangu-2.0-Pro weights and code. The edition also covers Nscale's Anyscale acquisition, Scale AI's appointment of Francis deSouza, NVIDIA's OO Agents framework, Jeff Dean's TPU origin story and Stony Brook research finding repeated sampling matched or beat seven tested Qwen reasoning methods at equal output-token cost.

• frontier models
• AI robotics
• AI infrastructure
• open-weight models
• AI agents
• AI research
• AI business
• AI policy

Stories in this roundup

1. 01
   OpenAI cuts GPT-5.6 Luna prices by 80% and Terra by 20%, and adds a Sol API mode offering up to 2.5× the speed at twice the standard priceOpenAI / X
2. 02
   The EU opens tenders for up to seven AI gigafactories, pairing €10 billion in public funding with at least €20 billion in expected private investmentEuropean Commission
3. 03
   Google DeepMind launches Gemini Robotics ER 2 for real-time video progress tracking, tool orchestration and multi-robot collaborationGoogle DeepMind
4. 04
   DeepSeek puts V4 Flash into public beta with native Responses API and Codex support, claiming an 82.7 Terminal-Bench 2.1 score after further post-trainingDeepSeek / X
5. 05
   Nscale agrees to acquire Ray creator Anyscale, adding its distributed-AI software layer to Nscale's data centres, GPUs and cloud platformNscale
6. 06
   Citadel buys most of Situational Awareness's public stock portfolio after steep losses force Leopold Aschenbrenner's AI-focused fund to unwind positionsFinancial Times
7. 07
   Scale AI appoints former Google Cloud COO and Illumina chief Francis deSouza as CEO, effective August 10Scale AI
8. 08
   MiniMax launches H3, an omni-modal model generating 15-second 2K video with native stereo audio, with open weights promised within daysMiniMax
9. 09
   Thinking Machines releases Inkling-Small, a 276B-parameter open-weight MoE with 12B active and native reasoning across text, images and audioThinking Machines
10. 10
    Huawei releases openPangu-2.0-Pro's 505B-parameter weights, inference code and technical report for Ascend hardwareHuawei
11. 11
    Jeff Dean tells YC's Diana Hu how napkin math sparked the TPU, why specialized inference chips have room to improve and where small AI teams can still winY Combinator / YouTube
12. 12
    NVIDIA releases research-preview OO Agents, a Python framework that expresses agent state, capabilities, prompts and typed contracts through classesNVIDIA / GitHub
13. 13
    Stony Brook PhD student Iliya Mirzaei finds none of seven test-time reasoning methods beat repeated sampling at equal output-token cost on Qwen2.5 math testsStony Brook University / arXiv