Call Params

cachedCall uses the same call shape, except reserveUsage and refundUsage belong at the cachedCall top level. Its top-level signal controls that caller's cancellation lifecycle. jsonMode/schema selection works the same way for cachedCall's return type as it does for call.

Pass null to either field to explicitly skip an instance-level defaultReasoningEffort/ defaultBudgetTokens for one call, the same way temperature: null opts a call out of defaultTemperature above. Omitting the field entirely (undefined) defers to the instance default instead, null is the only way to say "not for this call" once an instance default is set. Useful for a call using a forced toolChoice, since Anthropic rejects any reasoning alongside one, see the warning below.

Anthropic rejects budgetTokens/reasoningEffort combined with a toolChoice that forces tool use (a specific tool, or 'required'), a call that tries throws LLMError('invalid_params') before any request is sent, both on fromAnthropic and on Claude models called through fromBedrock. This also applies implicitly whenever jsonSchema forces a single synthetic tool call to emulate structured output on a model without native support, even with no toolChoice of your own set. Use toolChoice: 'auto' (or omit it), or pass budgetTokens: null/ reasoningEffort: null for that specific call.

Claude Opus 4.7 and later, Claude Opus 5, and every Claude 5 tier model (Sonnet 5, Fable 5, Mythos 5) reject temperature (and top_p/top_k) on a non-default value with a 400 error. Omit temperature entirely for those models, or pass null, and describe the desired behavior in the prompt instead.

Separately, Anthropic rejects temperature alongside any thinking mode, manual or adaptive, on every model that supports thinking at all, not just the models above. fromAnthropic and fromBedrock both drop temperature from the request automatically whenever budgetTokens or reasoningEffort is set, so you don't need to handle this yourself, it's called out here so the omission isn't a surprise if you're inspecting the request.

Claude Opus 4.7 and later, and every Claude 5 tier model, also reject manual, budget-based thinking (thinking: { type: 'enabled', budget_tokens }) outright, with a 400 error, separate from the temperature restriction above. fromAnthropic and fromBedrock detect these models and send adaptive thinking plus Anthropic's own effort parameter (low/medium/high/xhigh/ max) instead, converting reasoningEffort/budgetTokens onto the nearest effort level automatically. xhigh and max have no equivalent in VernLLM's four-tier reasoningEffort and aren't reachable through this conversion.

Gemini 3 series models use a different control, thinkingConfig.thinkingLevel (MINIMAL/LOW/MEDIUM/HIGH), not thinkingBudget. fromGemini detects Gemini 3 and later models (a version threshold, so 3.1, 3.5, and every future 3.x release are covered automatically) and sends thinkingLevel there instead. reasoningEffort maps onto it directly, no conversion table needed, VernLLM's four tiers line up exactly with Gemini's own four levels. budgetTokens alone is converted to the nearest tier first, through the same table used everywhere else, then mapped onto thinkingLevel. 0 and -1 have no thinkingLevel equivalent and both collapse to MINIMAL, the closest available approximation of "off", which several Gemini 3 models can't be fully disabled on anyway, MINIMAL still requires thought signatures rather than genuinely turning thinking off, unlike thinkingBudget: 0 on Gemini 2.5. Pass thinkingLevelModels as part of fromGemini's second argument to mark an additional model as using thinkingLevel, additive over the built-in threshold, same shape as reasoningEffortTokens above.