Rosetta Intel

OpenAI / Anthropic / Google product and model updates · 🛡 = security-relevant

OpenAI

GPT-5.6 preview(Sol / Terra / Luna): opened to a small set of government-vetted organizations ahead of the broader July release.
OpenAI reportedly offered the U.S. government (Washington) a 5% equity stake, sparking discussion on governance and government-industry relations.

Claude Sonnet 5released.
Claude Science: moving into scientific research, building on the acquisition of Coefficient Bioand the hiring of John Jumper, AlphaFold team lead and 2024 Nobel laureate in Chemistry.
Interpretability progress: using a Jacobian-based technique, identified a small internal subspace within the model whose behavior resembles the "global workspace" of consciousness neuroscience; DeepMind's Neel Nanda contributed commentary.
People: Andrej Karpathy joined Anthropic, returning to R&D with a focus on frontier LLMs.
Policy: the U.S. Department of Commerce lifted the export controls that had previously taken Anthropic's Fable 5 / Mythos 5offline.

GPT-5.6's tiered rollout to "government-vetted organizations" continues the main storyline of controlled frontier-model releases and export-control maneuvering (contrast with Commerce lifting the restrictions on Fable 5 / Mythos 5).
Anthropic's Jacobian interpretability work has potential defensive value for auditing model internals / detecting backdoors and jailbreaks— worth continued tracking.

Sources: OpenAI News, Anthropic News/Research, Google DeepMind Blog, Google AI Blog (compiled via WebSearch)