diff --git a/specs/10-ai-assistant.md b/specs/10-ai-assistant.md
index e17c4898..221096c7 100644
--- a/specs/10-ai-assistant.md
+++ b/specs/10-ai-assistant.md
@@ -43,7 +43,7 @@ never reach clients). The assistant acts on documents through the **same tool la
   | Fast | `fastModel` | `low` |
   Defaults per provider (`AiModels`, the only place with model ids; `ai.models.<provider>`
   overrides them): Anthropic `claude-opus-5-5` / `claude-haiku-5-5`, OpenAI `gpt-5` /
-  `gpt-5-mini`, Gemini `gemini-2.5-pro` / `gemini-2.5-flash`. Inline completions use the fast
+  `gpt-5-mini`, Gemini `gemini-3.1-pro-preview` / `gemini-3.8-flash` (the 2.5 models return 404 for new keys). Inline completions use the fast
   model. Effort is sent to OpenAI as `reasoning_effort`; dartantic has no effort option for
   Anthropic and Gemini, which use their model defaults. No `temperature`/`top_p`.
 - **Streaming** for every request; `maxTokens` 64000 (chat), 32000 (inline/proofread),
diff --git a/wonky_server/config/development.yaml b/wonky_server/config/development.yaml
index 50efa3c4..28844c73 100644
--- a/wonky_server/config/development.yaml
+++ b/wonky_server/config/development.yaml
@@ -127,7 +127,7 @@ wonky:
     # models:
     #   anthropic: {model: claude-opus-5-5, fastModel: claude-haiku-5-5}
     #   openai: {model: gpt-5, fastModel: gpt-5-mini}
-    #   gemini: {model: gemini-2.5-pro, fastModel: gemini-2.5-flash}
+    #   gemini: {model: gemini-3.1-pro-preview, fastModel: gemini-3.8-flash}
     dailyTokenBudget: 2000000
     maxConcurrentRequests: 2
     requestTimeoutMinutes: 10
diff --git a/wonky_server/lib/src/platform/ai/ai_config.dart b/wonky_server/lib/src/platform/ai/ai_config.dart
index 7f87bbf2..08043892 100644
--- a/wonky_server/lib/src/platform/ai/ai_config.dart
+++ b/wonky_server/lib/src/platform/ai/ai_config.dart
@@ -35,7 +35,7 @@ class AiProviderModels {
 ///     models:                       # optional; defaults in AiModels
 ///       anthropic: {model: claude-opus-5-5, fastModel: claude-haiku-5-5}
 ///       openai: {model: gpt-5, fastModel: gpt-5-mini}
-///       gemini: {model: gemini-2.5-pro, fastModel: gemini-2.5-flash}
+///       gemini: {model: gemini-3.1-pro-preview, fastModel: gemini-3.8-flash}
 ///     dailyTokenBudget: 2000000     # per user, all models
 ///     maxConcurrentRequests: 2      # per user
 ///     requestTimeoutMinutes: 10
diff --git a/wonky_server/lib/src/platform/ai/ai_models.dart b/wonky_server/lib/src/platform/ai/ai_models.dart
index 6faa85f5..dc463007 100644
--- a/wonky_server/lib/src/platform/ai/ai_models.dart
+++ b/wonky_server/lib/src/platform/ai/ai_models.dart
@@ -15,8 +15,8 @@ abstract final class AiModels {
 
   static const gpt = 'gpt-5';
   static const gptMini = 'gpt-5-mini';
-  static const geminiPro = 'gemini-2.5-pro';
-  static const geminiFlash = 'gemini-2.5-flash';
+  static const geminiPro = 'gemini-3.1-pro-preview';
+  static const geminiFlash = 'gemini-3.8-flash';
 
   /// The default `(model, fastModel)` of [provider].
   static (String, String) defaults(AiProvider provider) => switch (provider) {