What’s New
See Anthropic’s model migration guide for a full overview of changes. Claude 5 Sonnet introduces three major changes versus Sonnet 4.6:- Sampling parameters removed —
temperature,top_p, andtop_kare no longer supported and will be ignored - Adaptive-only thinking — when reasoning is enabled the only supported mode is adaptive;
thinking.budget_tokensis no longer supported andreasoning.max_tokensis ignored - New
'xhigh'effort level — a new effort level between'high'and'max'viaverbosity/output_config.effort
Sampling Parameters Removed
Claude 5 Sonnet no longer acceptstemperature, top_p, or top_k. If you pass these parameters, they will be silently ignored — your request will still succeed, but the parameters will have no effect.
Adaptive-Only Thinking
Claude 5 Sonnet supports only adaptive thinking. On Sonnet 4.6, reasoning could be controlled via a token budget (reasoning.max_tokens / thinking.budget_tokens) or left adaptive; on Sonnet 5, budget-based thinking is removed and adaptive is the only remaining mode when reasoning is on. The new API default is adaptive thinking on at effort high.
Reasoning itself remains opt-in on all Anthropic models via reasoning.enabled=true — Sonnet 5 does not change that.
Concretely on Sonnet 5:
reasoning.max_tokensis accepted but ignoredreasoning.effortmaps to Anthropic’soutput_config.effort(see Parameter Summary) rather than a thinking budgetthinking.budget_tokensis no longer supported upstream
verbosity. It maps to Anthropic’s output_config.effort and applies whether or not reasoning is enabled.
New 'xhigh' Effort Level
A new 'xhigh' effort level is available between 'high' and 'max' via the verbosity parameter. This maps to Anthropic’s output_config.effort.
low → medium → high → xhigh → max.
Sonnet 4.6 supported
low, medium, high, and max. Sonnet 5 adds 'xhigh' between 'high' and 'max'. For models that don’t support a given level, it automatically falls back to the next supported level ('xhigh' → 'high').Cyber Safeguards
Like Opus 4.7 and 4.8, Claude 5 Sonnet ships with real-time cyber classifiers that block some high-risk dual-use activities. Affected requests are refused upstream by Anthropic; OpenRouter surfaces the refusal as-is. See Anthropic’s Real-time cyber safeguards on Claude for details.Parameter Summary
With sampling parameters and reasoning budgets removed on Sonnet 5,output_config.effort is the remaining lever for influencing overall response effort. You can set it via verbosity, or via reasoning.effort when reasoning is enabled (verbosity wins if both are passed; 'minimal' maps to 'low', and 'none' disables reasoning entirely so no output_config.effort is sent).