Two items from this week warrant a decision. Everything else is tracking.
Grok 4.6
Released 12 August by SpaceXAI. It is a post-training upgrade on the 4.5 base.
The composite index is a four-way tie
Opus 5 scores 63, Fable 5 62, Grok 4.6 and GPT-5.6 Sol Max both 61.
The per-benchmark split is where the signal is
Grok 4.6 leads the knowledge-work evals and trails on software engineering.
EU AI Act, Article 50
Transparency obligations took effect 2 August 2026.
Vendor response, and its limits
Anthropic will watermark Claude text output worldwide.
What this asks of us
Decide whether watermarked output is acceptable in our own deliverables, and say so in policy before someone discovers the answer by accident.