From a498f28d144567c561eafeff1b6715391e8ee2a5 Mon Sep 17 00:00:00 2001
From: Enrico Ros <enrico.ros@gmail.com>
Date: Fri, 6 Feb 2026 14:59:08 -0800
Subject: [PATCH] LLMs: Anthropic: support for max effort. #962

---
 src/common/stores/llms/llms.parameters.ts                 | 8 ++++++++
 src/modules/aix/client/aix.client.ts                      | 4 ++--
 src/modules/aix/server/api/aix.wiretypes.ts               | 2 +-
 .../aix/server/dispatch/wiretypes/anthropic.wiretypes.ts  | 2 +-
 src/modules/llms/server/anthropic/anthropic.models.ts     | 4 ++--
 src/modules/llms/server/llm.server.types.ts               | 1 +
 6 files changed, 15 insertions(+), 6 deletions(-)

diff --git a/src/common/stores/llms/llms.parameters.ts b/src/common/stores/llms/llms.parameters.ts
index cb6484b1d..1cabb6a31 100644
--- a/src/common/stores/llms/llms.parameters.ts
+++ b/src/common/stores/llms/llms.parameters.ts
@@ -136,6 +136,14 @@ export const DModelParameterRegistry = {
     // No initialValue - undefined means off (e.g. default 200K context window)
   },
 
+  llmVndAntEffortMax: { // introduced with Claude Opus 4.6; this adds the 'max' level on top of llmVndAntEffort
+    label: 'Effort',
+    type: 'enum' as const,
+    description: 'Controls thinking depth. max = deepest reasoning with no constraints, high = default.',
+    values: ['low', 'medium', 'high', 'max'] as const,
+    // No initialValue - undefined means high effort (default)
+  } as const,
+
   llmVndAntEffort: {
     label: 'Effort',
     type: 'enum' as const,
diff --git a/src/modules/aix/client/aix.client.ts b/src/modules/aix/client/aix.client.ts
index c3bfdd800..4004e46b8 100644
--- a/src/modules/aix/client/aix.client.ts
+++ b/src/modules/aix/client/aix.client.ts
@@ -49,7 +49,7 @@ export function aixCreateModelFromLLMOptions(
   // destructure input with the overrides
   const {
     llmRef, llmTemperature, llmResponseTokens, llmTopP, llmForceNoStream,
-    llmVndAnt1MContext, llmVndAntSkills, llmVndAntThinkingBudget, llmVndAntWebFetch, llmVndAntWebSearch, llmVndAntEffort,
+    llmVndAnt1MContext, llmVndAntSkills, llmVndAntThinkingBudget, llmVndAntWebFetch, llmVndAntWebSearch, llmVndAntEffort, llmVndAntEffortMax,
     llmVndGeminiAspectRatio, llmVndGeminiImageSize, llmVndGeminiCodeExecution, llmVndGeminiComputerUse, llmVndGeminiGoogleSearch, llmVndGeminiMediaResolution, llmVndGeminiShowThoughts, llmVndGeminiThinkingBudget, llmVndGeminiThinkingLevel, llmVndGeminiThinkingLevel4,
     llmVndMoonReasoningEffort, // -> mapped to vndOaiReasoningEffort below
     // llmVndMoonshotWebSearch,
@@ -110,7 +110,7 @@ export function aixCreateModelFromLLMOptions(
     ...(llmVndAntSkills ? { vndAntSkills: llmVndAntSkills } : {}),
     ...(llmVndAntWebFetch === 'auto' ? { vndAntWebFetch: llmVndAntWebFetch } : {}),
     ...(llmVndAntWebSearch === 'auto' ? { vndAntWebSearch: llmVndAntWebSearch } : {}),
-    ...(llmVndAntEffort ? { vndAntEffort: llmVndAntEffort } : {}),
+    ...((llmVndAntEffortMax || llmVndAntEffort) ? { vndAntEffort: llmVndAntEffortMax || llmVndAntEffort } : {}),
     ...(llmVndGeminiAspectRatio ? { vndGeminiAspectRatio: llmVndGeminiAspectRatio } : {}),
     ...(llmVndGeminiCodeExecution === 'auto' ? { vndGeminiCodeExecution: llmVndGeminiCodeExecution } : {}),
     ...(llmVndGeminiComputerUse ? { vndGeminiComputerUse: llmVndGeminiComputerUse } : {}),
diff --git a/src/modules/aix/server/api/aix.wiretypes.ts b/src/modules/aix/server/api/aix.wiretypes.ts
index accff9d6a..310ebe0da 100644
--- a/src/modules/aix/server/api/aix.wiretypes.ts
+++ b/src/modules/aix/server/api/aix.wiretypes.ts
@@ -451,7 +451,7 @@ export namespace AixWire_API {
 
     // Anthropic
     vndAnt1MContext: z.boolean().optional(),
-    vndAntEffort: z.enum(['low', 'medium', 'high']).optional(),
+    vndAntEffort: z.enum(['low', 'medium', 'high', 'max']).optional(),
     vndAntSkills: z.string().optional(),
     vndAntThinkingBudget: z.number().or(z.literal('adaptive')).nullable().optional(),
     vndAntToolSearch: z.enum(['regex', 'bm25']).optional(), // Tool Search Tool variant
diff --git a/src/modules/aix/server/dispatch/wiretypes/anthropic.wiretypes.ts b/src/modules/aix/server/dispatch/wiretypes/anthropic.wiretypes.ts
index c48eef7a4..e30f59216 100644
--- a/src/modules/aix/server/dispatch/wiretypes/anthropic.wiretypes.ts
+++ b/src/modules/aix/server/dispatch/wiretypes/anthropic.wiretypes.ts
@@ -847,7 +847,7 @@ export namespace AnthropicWire_API_Message_Create {
      * Allows trading off response thoroughness for efficiency (Claude Opus 4.5+ only).
      */
     output_config: z.object({
-      effort: z.enum(['low', 'medium', 'high']).optional(),
+      effort: z.enum(['low', 'medium', 'high', 'max']).optional(),
     }).optional(),
 
     /**
diff --git a/src/modules/llms/server/anthropic/anthropic.models.ts b/src/modules/llms/server/anthropic/anthropic.models.ts
index 677a24bb4..98ce32041 100644
--- a/src/modules/llms/server/anthropic/anthropic.models.ts
+++ b/src/modules/llms/server/anthropic/anthropic.models.ts
@@ -40,7 +40,7 @@ const _hardcodedAnthropicVariants: ModelVariantMap = {
     label: 'Claude Opus 4.6 (Thinking)',
     description: 'Claude Opus 4.6 with adaptive thinking mode for the most complex reasoning and agentic workflows',
     interfaces: [...IF_4_R, LLM_IF_ANT_ToolsSearch],
-    parameterSpecs: [...ANT_TOOLS, { paramId: 'llmVndAntThinkingBudget', hidden: true, initialValue: -1 /* adaptive */ }, { paramId: 'llmVndAntEffort' }, { paramId: 'llmVndAnt1MContext' }],
+    parameterSpecs: [...ANT_TOOLS, { paramId: 'llmVndAntThinkingBudget', hidden: true, initialValue: -1 /* adaptive */ }, { paramId: 'llmVndAntEffortMax' }, { paramId: 'llmVndAnt1MContext' }],
     // benchmark: { cbaElo: ... }, // TBD
     maxCompletionTokens: 32000,
   },
@@ -136,7 +136,7 @@ export const hardcodedAnthropicModels: (ModelDescriptionSchema & { isLegacy?: bo
     contextWindow: 200000,
     maxCompletionTokens: 128000,
     interfaces: [...IF_4, LLM_IF_ANT_ToolsSearch],
-    parameterSpecs: [...ANT_TOOLS, { paramId: 'llmVndAntEffort' }, { paramId: 'llmVndAnt1MContext' }],
+    parameterSpecs: [...ANT_TOOLS, { paramId: 'llmVndAntEffortMax' }, { paramId: 'llmVndAnt1MContext' }],
     // Note: Tiered pricing - ≤200K: $5/$25, >200K: $10/$37.50 (with 1M context enabled)
     // Cache pricing also tiered: write 1.25× input, read 0.10× input
     chatPrice: {
diff --git a/src/modules/llms/server/llm.server.types.ts b/src/modules/llms/server/llm.server.types.ts
index 8dce721e5..88ced8f68 100644
--- a/src/modules/llms/server/llm.server.types.ts
+++ b/src/modules/llms/server/llm.server.types.ts
@@ -79,6 +79,7 @@ const ModelParameterSpec_schema = z.object({
     // Anthropic
     'llmVndAnt1MContext',
     'llmVndAntEffort',
+    'llmVndAntEffortMax',
     'llmVndAntSkills',
     'llmVndAntThinkingBudget',
     'llmVndAntWebFetch',