Claude: map Reasoning Effort to adaptive thinking config (#5219)

Supersedes #5105

public/index.html

@@ -2182,6 +2182,9 @@
                                             <div class="toggle-description justifyLeft marginBot5" data-i18n="Constrains the verbosity of the model's response.">
                                                 Constrains the verbosity of the model's response.
                                             </div>
+                                            <strong class="toggle-description justifyLeft marginBot5" data-source="claude">
+                                                On Opus 4.6 / Sonnet 4.6, a non-automatic Reasoning Effort takes precedence over Verbosity.
+                                            </strong>
                                         </div>
                                     </div>
                                     <div class="range-block" data-source="claude">

src/endpoints/backends/chat-completions.js

@@ -228,6 +228,7 @@ async function sendClaudeRequest(request, response) {
         const isLimitedSampling = /^claude-(opus-4-1|sonnet-4-5|haiku-4-5|opus-4-5|opus-4-6|sonnet-4-6)/.test(request.body.model);
         const useVerbosity = /^claude-(opus-4-5|opus-4-6|sonnet-4-6)/.test(request.body.model);
         const noPrefillModel = /^claude-(opus-4-6|sonnet-4-6)/.test(request.body.model);
+        const isAdaptiveModel = /^claude-(opus-4-6|sonnet-4-6)/.test(request.body.model);
         let fixThinkingPrefill = false;
         // Add custom stop sequences
         const stopSequences = [];
@@ -305,10 +306,18 @@ async function sendClaudeRequest(request, response) {
         }
 
         const reasoningEffort = request.body.reasoning_effort;
-        const budgetTokens = calculateClaudeBudgetTokens(requestBody.max_tokens, reasoningEffort, requestBody.stream);
+        const budgetTokens = calculateClaudeBudgetTokens(requestBody.max_tokens, reasoningEffort, requestBody.stream, isAdaptiveModel);
 
-        if (useThinking && Number.isInteger(budgetTokens)) {
-            // No prefill when thinking
+        // Adaptive thinking: returns a string effort level (like Gemini 3)
+        if (useThinking && typeof budgetTokens === 'string') {
+            fixThinkingPrefill = true;
+            requestBody.thinking = { type: 'adaptive' };
+            requestBody.output_config ??= {};
+            requestBody.output_config.effort = budgetTokens;
+            // top_k is not allowed in adaptive mode
+            delete requestBody.top_k;
+        } else if (useThinking && Number.isInteger(budgetTokens)) {
+            // Traditional thinking: returns a numeric budget
             fixThinkingPrefill = true;
             const minThinkTokens = 1024;
             if (requestBody.max_tokens <= minThinkTokens) {
@@ -332,8 +341,8 @@ async function sendClaudeRequest(request, response) {
             convertedPrompt.messages[convertedPrompt.messages.length - 1].role = 'user';
         }
 
-        // Verbosity = 'effort' (same values as OpenAI)
-        if (useVerbosity && request.body.verbosity) {
+        // Verbosity = 'effort' (same values as OpenAI) - only if not already set by adaptive thinking
+        if (useVerbosity && request.body.verbosity && !requestBody.output_config?.effort) {
             betaHeaders.push('effort-2025-11-24');
             requestBody.output_config ??= {};
             requestBody.output_config.effort = request.body.verbosity;

src/prompt-converters.js

@@ -1110,12 +1110,33 @@ export function cachingSystemPromptForOpenRouter(messages, ttl = undefined) {
 
 /**
  * Calculate the Claude budget tokens for a given reasoning effort.
+ * Returns a string effort level for adaptive thinking (Opus 4.6+), a number for traditional thinking, or null for auto.
  * @param {number} maxTokens Maximum tokens
  * @param {string} reasoningEffort Reasoning effort
  * @param {boolean} stream If streaming is enabled
- * @returns {number?} Budget tokens
+ * @param {boolean} isAdaptiveModel If the model supports adaptive thinking (Opus 4.6+)
+ * @returns {number|string|null} Budget tokens, effort string, or null
  */
-export function calculateClaudeBudgetTokens(maxTokens, reasoningEffort, stream) {
+export function calculateClaudeBudgetTokens(maxTokens, reasoningEffort, stream, isAdaptiveModel) {
+    // Adaptive thinking for Opus 4.6+: return effort string (like Gemini 3)
+    if (isAdaptiveModel) {
+        switch (reasoningEffort) {
+            case REASONING_EFFORT.auto:
+                return null;
+            case REASONING_EFFORT.min:
+                return 'low';
+            case REASONING_EFFORT.low:
+                return 'low';
+            case REASONING_EFFORT.medium:
+                return 'medium';
+            case REASONING_EFFORT.high:
+                return 'high';
+            case REASONING_EFFORT.max:
+                return 'max';
+        }
+        return null;
+    }
+
     let budgetTokens = 0;
 
     switch (reasoningEffort) {