AI Daily — July 30, 2026
2026-07-30
openai.com
How GPT-5.6 fuses frontier intelligence with frontier efficiency
OpenAI has published a technical overview of GPT-5.6, focusing on efficiency improvements across model inference and agentic workflows to deliver higher capability per compute dollar. The release targets cost-effectiveness at scale rather than raw benchmark maximums, signaling a shift in optimization priorities for production deployments.
openai.com
How enabling two settings tripled our scores on the ARC-AGI-3 benchmark
OpenAI reports that enabling reasoning retention and context compaction in the GPT-5.6 API tripled its scores on ARC-AGI-3 without architectural changes. The findings highlight how inference-time configuration — specifically preserving reasoning state across turns and compacting long contexts — can dramatically affect benchmark performance on abstract reasoning tasks.
openai.com
OpenAI gives 100,000 academic researchers free access to advanced ChatGPT models
OpenAI is providing 100,000 academic researchers with free access to its most capable ChatGPT models to accelerate scientific discovery and collaboration. The program represents a significant expansion of API access for the research community, potentially lowering barriers to AI-assisted research workflows at scale.
deepmind.google
Google DeepMind launches Lyria 3.5 in Google Flow Music
Google DeepMind has released Lyria 3.5, an updated music generation model integrated into Google Flow Music, with improvements in musicality, lyric generation, vocal synthesis, and fine-grained creative control. The update represents a meaningful step forward in AI-generated audio quality and user-directed composition.
rss.arxiv.org
Probing the Origins of Reasoning Performance: RL vs. SFT Fine-Tuned Models
This paper provides mechanistic evidence for why RL-trained reasoning models outperform SFT counterparts on math tasks, using linear probes on hidden states and mean ablation studies. RL models develop more linearly separable internal representations and a hierarchical layer structure where deeper layers specialize in correctness-predictive features — offering interpretability insights beyond benchmark comparisons.