Fix GPT-5 max_completion_tokens parameter detection

GPT-5 and newer models require max_completion_tokens instead of max_tokens.
Added usesNewTokenParam() to detect GPT-5+ models separately from reasoning
model restrictions (temperature, json_object, system messages).

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

src/lib/openai.ts

@@ -19,8 +19,15 @@ const globalForOpenAI = globalThis as unknown as {
  */
 const REASONING_MODEL_PREFIXES = ['o1', 'o3', 'o4']
 
+/**
+ * Models that use max_completion_tokens instead of max_tokens.
+ * This includes reasoning models AND newer GPT models (GPT-5+).
+ */
+const NEW_TOKEN_PARAM_PREFIXES = ['o1', 'o3', 'o4', 'gpt-5', 'gpt-6', 'gpt-7']
+
 /**
  * Check if a model is a reasoning model (o1, o3, o4 series)
+ * These models have additional restrictions (no temperature, no json_object, etc.)
  */
 export function isReasoningModel(model: string): boolean {
   const modelLower = model.toLowerCase()
@@ -31,6 +38,19 @@ export function isReasoningModel(model: string): boolean {
   )
 }
 
+/**
+ * Check if a model requires max_completion_tokens instead of max_tokens.
+ * This includes reasoning models AND newer GPT models (GPT-5+).
+ */
+export function usesNewTokenParam(model: string): boolean {
+  const modelLower = model.toLowerCase()
+  return NEW_TOKEN_PARAM_PREFIXES.some(prefix =>
+    modelLower.startsWith(prefix) ||
+    modelLower.includes(`/${prefix}`) ||
+    modelLower.includes(`-${prefix}`)
+  )
+}
+
 // ─── Chat Completion Parameter Builder ───────────────────────────────────────
 
 type MessageRole = 'system' | 'user' | 'assistant' | 'developer'
@@ -84,8 +104,9 @@ export function buildCompletionParams(
   }
 
   // Token limit parameter differs between model types
+  // Newer models (GPT-5+, o-series) use max_completion_tokens
   if (options.maxTokens) {
-    if (isReasoning) {
+    if (usesNewTokenParam(model)) {
       params.max_completion_tokens = options.maxTokens
     } else {
       params.max_tokens = options.maxTokens