Claude Sonnet 5 vs Sonnet 4.6 prompt steering benchmark

If you’re weighing Sonnet 5 against Sonnet 4.6 for everyday Claude Code work, here is the short version. Sonnet 5 is the better default for most jobs. It answers more concisely, runs much faster, and is now permanently cheaper than 4.6 on every task in this benchmark. But it isn’t strictly better. Set the price discount aside and at high effort it’s the less efficient of the two, it follows precise instructions less reliably, and it reacts to “work harder” prompt phrasing in ways 4.6 shrugs off. Reach for 4.6 when you need exact rules followed on a budget.A few terms first, since the post leans on them.Prompt steering. Prepending a short fixed sentence to every prompt to nudge the model, like “be concise” or “think step by step”. I call each such sentence a wrapper.Effort. Claude Code’s reasoning dial, low, medium, high, max on Sonnet. More effort means more reasoning, usually better answers and more tokens.Pricing. Sonnet 5 runs at $2/$10 per million input/output tokens, a third…