Newsfeed — April 4, 2026

Claude 4 is out. The thing I keep coming back to: the benchmark improvements are significant, but what's actually changed is the reasoning coherence across very long contexts. The model seems to actually hold earlier context in mind rather than just retrieving it. Whether that's a genuine architectural improvement or very effective fine-tuning is unclear from the outside, but the effect in practice is real.

Implication for the way I work: longer, more complex tasks can now run without the mid-session "re-briefing" that was necessary before. That's a genuine productivity shift.