Grok 5 arrived as the largest publicly announced model, a mixture-of-experts design with 6 trillion total parameters. For buyers the honest answer is that parameter counts stopped predicting usefulness: routing, training data and tool integration decide real-world results, and small specialised models keep winning on cost.
Six trillion parameters is a headline built for repetition, and it worked. But mixture-of-experts architectures only activate a fraction of those weights per request, so the number describes capacity, not the compute you actually get per answer.
Across 2026 the pattern held: benchmark gaps between frontier models narrowed while the price of good-enough intelligence collapsed. The buying question changed from which model is smartest to which system is most reliable inside your workflow.
A sane evaluation method
Collect twenty real tasks from your business, actual emails, actual documents, actual quotes. Run them through the candidates. Score outputs blind. The winner on your work beats the winner on someone else's leaderboard.
Re-run the same test quarterly. Releases now land monthly, and last quarter's loser is frequently this quarter's best value.