Models

A visual-reasoning study targets the information held between seeing an image and answering a question. Its training method makes those hidden states more sensitive to relevant image changes.

A controlled study separates who generates training answers from how a student model learns. Changing the objective and learning rate explains much of the difference.

Version 2.0 uses two teachers to preserve casual dialogue while improving instruction and tool behaviour. The release includes local weights and explicit tradeoffs.

SoftBank paid the final $10 billion tranche of its 2026 OpenAI commitment, bringing its cumulative investment to $64.6 billion.

Google has unveiled its largest Gemini model but has not opened it to the public. Independent reporting shows a deliberately narrow launch after delays and a cancelled intermediate model.

Anthropic says the open-weight model built working browser exploits and that common techniques weakened its safeguards. A US government assessment independently confirms a sharp capability gain.

OpenAI's new mid-tier model costs one-fifth as much as GPT-6 Astra at standard API rates. Its performance and safety comparisons remain company-run evaluations.