Specialized capability arrived with specialized data and access risks. OpenAI put GPT-Rosalind behind qualification, US officials pursued access to Anthropic's cyber model, and Gemini began drawing on private photo libraries for image generation. A video-reasoning study added a cost constraint: most gains from longer thought streams arrived within the first few hundred tokens.
1. GPT-Rosalind opens through a qualified research preview
OpenAI introduced GPT-Rosalind for biology, drug discovery, and translational medicine through a qualified US enterprise preview. A free Codex plugin connects mainline models to more than 50 public scientific tools and databases, while access to the specialized model follows a separate review.
The split preserves broad tool access while gating deeper biological reasoning. OpenAI reports strong public and partner evaluations, including best-of-ten submissions on an unpublished RNA task; that setup measures search with multiple attempts, not a single autonomous scientific decision. The model remains upstream of experiments and clinical validation.
Sources: OpenAI on GPT-Rosalind · Bloomberg on OpenAI's drug-discovery model
2. The White House pursues agency access to restricted Mythos
The White House was reportedly working to provide US agencies with access to Anthropic's restricted Mythos cybersecurity model. The proposal would place federal systems and contractors inside an access program originally designed to give selected defenders an early security window.
Government participation changes the disclosure chain. A flaw found in a vendor product can touch classification rules, procurement relationships, and civilian patching at once. The unresolved Pentagon dispute also gives access terms political leverage beyond ordinary model licensing.
Sources: Reuters on proposed US agency access to Mythos
3. Gemini uses labeled Google Photos as image references
Google expanded Gemini image generation so Personal Intelligence can select labeled people and pets from an opted-in Google Photos library. The Sources control reveals which image Gemini chose, and a user can substitute a different reference before regenerating.
The feature turns photo organization into generation context. Labels created for search now help synthesize likenesses of family members or bystanders who may never interact with Gemini. Google's statement that private Photos are excluded from direct model training addresses one data path, while generated depictions create a separate consent problem.
Sources: Google on personalized images in Gemini · Ars Technica on the Google Photos integration
4. Video reasoning gains flatten after a few hundred thought tokens
A preprint evaluated four Gemini 2.5 Flash and Flash Lite configurations on scenes drawn from 100 hours of video. Most quality gains arrived within the first few hundred thought tokens, and Flash Lite offered the paper's best balance between output quality and token use.
Tight budgets produced a revealing failure: final answers sometimes added content absent from the preceding reasoning trace, which the authors call compression-step hallucination. The finding makes reasoning budget a quality parameter, but longer traces quickly hit diminishing returns on this model family and task set.