Sonnet 5 Is Now the Default. Swapping to It Isn't a One-Liner.
A mid-tier Sonnet just edged the flagship on knowledge work — and became the default overnight. But three silent default changes turn a model-string swap into a production incident nobody sees until load.
NeuroX AI · July 10, 2026
Anthropic shipped Claude Sonnet 5 on June 30 and made it the default for every Free and Pro user the next day. The number that should stop you isn't the price — it's that a mid-tier Sonnet scored 1,618 on GDPval-AA v2 knowledge work, edging the flagship Opus 4.8's 1,615 (MarkTechPost). On SWE-bench Pro, Opus still leads (69.2% vs 63.2%), but the gap where most teams actually live narrowed to a rounding error.
Here's the part a benchmark table won't tell you: pointing your agent at Sonnet 5 is not a one-line model-string change.
Three defaults shifted silently. temperature, top_p, and top_k are now rejected — a request that ran clean on Sonnet 4.6 with a tuned sampling value returns a 400. Adaptive thinking is on by default (it was off on 4.6), so a loop with a tight max_tokens now spends budget on reasoning and truncates mid-answer. And the new tokenizer counts roughly 30% more tokens for identical text, quietly blowing token budgets and context-window math that were calibrated last week.
None of these throw at demo time. They surface in production, under load, in the loop you thought you'd already shipped. A model that upgrades itself underneath you is exactly why the discipline has to live below the model.