Daily AI RoundupsThursday, 30 July 2026

The 30 July 2026 Calathea roundup leads with OpenAI’s account of a 20% serving-cost reduction for GPT-5.6 Sol and its GPT-Red system outperforming human red-teamers on unseen prompt-injection scenarios. A Princeton-led study finds frontier agents completed the engineering but failed both open-ended AI research projects, while SpaceXAI launches Grok Voice Think Fast 2.0 at $0.08 per audio minute. Tencent open-sources AngelSpec, Baseten launches Helios for provider-independent inference, Shanghai AI Laboratory publishes Intern-S2-Mobius and Google adds Gemini voice dictation to macOS. The edition also covers Moonshot AI’s $3.5 billion Series F, Microsoft 365 Copilot, Ontology 1, a JAX-to-LLVM compiler, Reflection AI, ChatGPT nearing one billion weekly users and Harvey’s legal-AI research hub.

• frontier models
• AI security
• AI agents
• inference infrastructure
• voice AI
• open-source AI
• AI research
• developer tools

Stories in this roundup

1. 01
   OpenAI says GPT-5.6 Sol helped cut model-serving costs 20% and raise speculative-decoding efficiency by more than 15% by optimizing kernels, routing and its own draft modelOpenAI
2. 02
   Moonshot AI completes joint-stock restructuring and closes a $3.5B Series F at a $35B valuationChinaZhiDX / X
3. 03
   OpenAI’s GPT-Red beat human red-teamers 84% to 13% on unseen prompt-injection scenarios; adversarial training cut GPT-5.6 Sol’s failures sixfoldOpenAI / arXiv
4. 04
   Princeton-led shadow evaluations find frontier agents completed the engineering but failed both six-day open-ended AI research projectsPrinceton / arXiv
5. 05
   SpaceXAI launches Grok Voice Think Fast 2.0 at $0.08 per audio minute, claiming 0.70-second time-to-first-audio and stronger agentic voice performanceSpaceXAI / X
6. 06
   Baseten launches Helios to spread inference traffic across cloud providers and switch regions or models in under two minutes when failures hitBaseten / X
7. 07
   Tencent open-sources AngelSpec for speculative decoding, claiming a 1.98–2.40× speedup for Hy3-A21B across tested concurrency levelsTencent Hunyuan / X
8. 08
   Microsoft launches a unified Microsoft 365 Copilot app with multi-agent workflows, Agent 365 management and PowerPoint generation from promptsSatya Nadella / X
9. 09
   Shanghai AI Laboratory’s InternLM team publishes Intern-S2-Mobius, the architecture behind a 397B scientific model that separates memory from reasoningInternLM / Hugging Face
10. 10
    Onton unveils Ontology 1, a neurosymbolic shopping-search model that beat Google and Amazon on its company-authored 90-query benchmarkOnton / X
11. 11
    Google adds natural-language voice dictation to Gemini on macOS, inserting cleaned-up transcripts, edits and summaries into the active appGoogle
12. 12
    Vercantez maps Cloudflare Durable Objects into a serverless agent architecture with isolated compute, durable state and hibernationVercantez / X
13. 13
    Xanadu’s Josh Iza explains how Catalyst accidentally became a JAX-to-LLVM compiler, bypassing XLA while retaining MLIR extensibility and Enzyme backpropagationJosh Iza
14. 14
    Nvidia-backed Reflection AI has yet to release a model nearly a year after raising $800M to build a US open-source AI contender, The Information reportsThe Information / X
15. 15
    ChatGPT nears 1B weekly active users seven months after OpenAI’s target, The Information reportsThe Information / X
16. 16
    Andrew Ho argues the next AI systems will be asynchronous, dynamic and distributed rather than single-model chat loopsAndrew Ho / X
17. 17
    Harvey launches a legal-AI research hub after open-sourcing its 1,200-task Legal Agent Benchmark across more than 24 practice areasHarvey / X