AI Daily — July 1, 2026
2026-07-01
www.anthropic.com
Introducing Claude Sonnet 5
Anthropic released Claude Sonnet 5, a new flagship model positioned as a major capability upgrade in the Sonnet line. The announcement coincides with a cluster of Anthropic product launches, suggesting a coordinated platform push.
www.technologyreview.com
Claude Science: Anthropic's autonomous research agent for scientific discovery
Anthropic announced Claude Science at an event targeting pharmaceutical and biotech researchers, positioning it as the scientific equivalent of Claude Code — capable of autonomously executing complex research tasks from high-level instructions. The product has access to tools and domain-specific resources relevant to life sciences workflows, mirroring the agentic coding paradigm applied to experimental science.
openai.com
OpenAI launches GeneBench-Pro genomics benchmark
OpenAI introduced GeneBench-Pro, a benchmark designed to evaluate AI performance on genomics, biology, and scientific research tasks using complex, real-world datasets. This signals OpenAI's push into rigorous scientific evals alongside its broader life-sciences ambitions.
deepmind.google
Google DeepMind releases Nano Banana 2 Lite and Gemini Omni Flash
Google DeepMind launched two new models: Nano Banana 2 Lite, a compact on-device model, and Gemini Omni Flash, a fast multimodal variant aimed at low-latency developer use cases. The releases extend the Gemini family toward efficiency-optimized deployment scenarios.
rss.arxiv.org
What Drives Interactive Improvement from Feedback? (arXiv)
Researchers introduce a controlled student-teacher protocol to isolate when natural-language feedback actually drives improvement in multi-turn language agent settings versus gains from mere resampling or additional compute. Evaluated across 13 open-weight models on Omni-MATH, Codeforces, BBEH Linguini, and ARC-AGI1, the study finds multi-turn gains are frequently attributable to resampling effects rather than semantically useful feedback. The methodology provides a cleaner framework for evaluating agent self-refinement and teacher-guided improvement.