diff --git a/src/lib/llm-providers.test.ts b/src/lib/llm-providers.test.ts
index f1bb2085a..c08b9bc92 100644
--- a/src/lib/llm-providers.test.ts
+++ b/src/lib/llm-providers.test.ts
@@ -518,6 +518,53 @@ describe("Azure OpenAI provider", () => {
     expect(body.temperature).toBeUndefined()
   })
 
+  it("auto-detects GPT-5 from an Azure deployment name that does not start with the model id", () => {
+    const cfg = getProviderConfig(mkConfig({
+      provider: "azure",
+      model: "my-gpt-5-chat",
+      customEndpoint: "https://example-resource.openai.azure.com",
+    }))
+    const body = cfg.buildBody(
+      [{ role: "user", content: "hi" }],
+      { max_tokens: 2048 },
+    ) as Record<string, unknown>
+
+    expect(body.max_tokens).toBeUndefined()
+    expect(body.max_completion_tokens).toBe(2048)
+  })
+
+  it("auto-detects the dashless gpt5 spelling and o-series names in Azure deployments", () => {
+    for (const model of ["prod-gpt5", "o3-reasoning"]) {
+      const cfg = getProviderConfig(mkConfig({
+        provider: "azure",
+        model,
+        customEndpoint: "https://example-resource.openai.azure.com",
+      }))
+      const body = cfg.buildBody(
+        [{ role: "user", content: "hi" }],
+        { max_tokens: 1000 },
+      ) as Record<string, unknown>
+
+      expect(body.max_tokens, model).toBeUndefined()
+      expect(body.max_completion_tokens, model).toBe(1000)
+    }
+  })
+
+  it("keeps max_tokens for non-GPT-5 Azure deployments", () => {
+    const cfg = getProviderConfig(mkConfig({
+      provider: "azure",
+      model: "gpt-4o-chat",
+      customEndpoint: "https://example-resource.openai.azure.com",
+    }))
+    const body = cfg.buildBody(
+      [{ role: "user", content: "hi" }],
+      { max_tokens: 3000 },
+    ) as Record<string, unknown>
+
+    expect(body.max_tokens).toBe(3000)
+    expect(body.max_completion_tokens).toBeUndefined()
+  })
+
   it("applies explicit Azure model family to custom Azure endpoints", () => {
     const cfg = getProviderConfig(mkConfig({
       provider: "custom",
diff --git a/src/lib/llm-providers.ts b/src/lib/llm-providers.ts
index ad7805259..c30b16694 100644
--- a/src/lib/llm-providers.ts
+++ b/src/lib/llm-providers.ts
@@ -376,13 +376,33 @@ function isGlmVisionModel(model: string): boolean {
     || /(?:^|[-_.])glm[-_.]4v(?:[-_.]|$)/i.test(normalized)
 }
 
+// GPT-5 / o-series (o1, o3, …) reasoning models require
+// `max_completion_tokens` instead of `max_tokens`. Azure deployment names are
+// user-chosen and frequently don't *start* with the underlying model id
+// (e.g. "prod-gpt5", "my-gpt-5-chat", "o3-reasoning"), so match a family token
+// anywhere in the name using word boundaries. `gpt-?5` also covers the common
+// dashless "gpt5" spelling. See issue #610.
+function deploymentNameSuggestsStrictCompletion(name: string): boolean {
+  const n = name.trim().toLowerCase()
+  return /(?:^|[-_.])gpt-?5(?:[-_.]|$)/.test(n) || /(?:^|[-_.])o\d+(?:[-_.]|$)/.test(n)
+}
+
 function isOpenAiStrictCompletionModel(config: LlmConfig): boolean {
-  if ((config.provider === "azure" || (config.provider === "custom" && isAzureOpenAiEndpoint(config.customEndpoint)))
-    && config.azureModelFamily === "gpt5") {
-    return true
-  }
+  const isAzureLike =
+    config.provider === "azure" ||
+    (config.provider === "custom" && isAzureOpenAiEndpoint(config.customEndpoint))
+
+  // Explicit family override always wins.
+  if (isAzureLike && config.azureModelFamily === "gpt5") return true
 
   const model = config.model.trim().toLowerCase()
+
+  // When the Azure family is left on "auto" (the default), infer it from the
+  // deployment name with a boundary-aware match anywhere in the string so
+  // GPT-5 / o-series deployments get `max_completion_tokens` without requiring
+  // the user to manually set the family. See issue #610.
+  if (isAzureLike && deploymentNameSuggestsStrictCompletion(model)) return true
+
   const strictModel =
     /^gpt-5(?:[.\-_]|$)/.test(model) || /^o\d+(?:[.\-_]|$)/.test(model)
   if (!strictModel) return false