Add argrouter submission

llm_inference/model_inference.py

@@ -12,6 +12,7 @@
 import logging
 from typing import Dict, Any
 from openai import OpenAI
+import re
 import tiktoken
 
 logger = logging.getLogger(__name__)
@@ -194,6 +195,11 @@ def _get_provider(self, model_name: str) -> str:
             "qwen/qwen3-next-80b-a3b-instruct": "openrouter",
             "Qwen/Qwen3-Coder-Next": "openrouter",
             "deepseek/deepseek-v4-flash": "openrouter",
+            "google/gemini-3.8-flash-reasoning-low": "openrouter",
+            "z-ai/glm-5.3-flash": "openrouter",
+            "openai/gpt-6-luna": "openrouter",
+            "xiaomi/mimo-v2.6-flash": "openrouter",
+            "google/gemma-4-31b-it": "openrouter",
             "x-ai/grok-4.1-fast": "openrouter",
             "mistralai/devstral-2512:free": "openrouter",
             "meta-llama/llama-3.3-70b-instruct": "openrouter",
@@ -321,8 +327,20 @@ def _call_openrouter(self, model_name: str, prompt: str) -> Dict[str, Any]:
             api_key=openrouter_api_key,
         )
 
+        # "<model>-reasoning-<effort>" selects the same model with an explicit
+        # OpenRouter reasoning effort (e.g. google/gemini-3.8-flash-reasoning-low).
+        api_model, extra_body = model_name, None
+        match = re.fullmatch(
+            r"(.+)-reasoning-(none|minimal|low|medium|high)", model_name
+        )
+        if match:
+            api_model = match.group(1)
+            extra_body = {"reasoning": {"effort": match.group(2)}}
+
         response = client.chat.completions.create(
-            model=model_name, messages=[{"role": "user", "content": prompt}]
+            model=api_model,
+            messages=[{"role": "user", "content": prompt}],
+            extra_body=extra_body,
         )
 
         usage = getattr(response, "usage", None)

model_cost/model_cost.json

@@ -386,5 +386,21 @@
   "google/gemma-4-31b-it": {
     "input_token_price_per_million": 0.08,
     "output_token_price_per_million": 0.35
+  },
+  "google/gemini-3.8-flash-reasoning-low": {
+    "input_token_price_per_million": 0.75,
+    "output_token_price_per_million": 3.75
+  },
+  "z-ai/glm-5.3-flash": {
+    "input_token_price_per_million": 0.15,
+    "output_token_price_per_million": 0.5
+  },
+  "openai/gpt-6-luna": {
+    "input_token_price_per_million": 0.1,
+    "output_token_price_per_million": 0.5
+  },
+  "xiaomi/mimo-v2.6-flash": {
+    "input_token_price_per_million": 0.14,
+    "output_token_price_per_million": 0.28
   }
 }

router_inference/config/argrouter.json

@@ -0,0 +1,14 @@
+{
+  "pipeline_params": {
+    "router_name": "argrouter",
+    "router_cls_name": "external_prediction_file",
+    "models": [
+      "google/gemini-3.8-flash-reasoning-low",
+      "google/gemma-4-31b-it",
+      "z-ai/glm-5.3-flash",
+      "openai/gpt-6-luna",
+      "xiaomi/mimo-v2.6-flash"
+    ],
+    "description": "Routes on the question content only: the first and last paragraph of the prompt (instruction and answer-format text) are dropped and field labels / option letters stripped with generic patterns; no RouterArena config, template or dataset file is read. Per-model P(correct) from logistic regression + kNN on local bge-small embeddings of that content, expected cost from the billed token usage (reasoning included) of the nearest external training prompts; picks argmax P(correct) - lambda * expected cost. Calibration sources weighted equally. Trained only on held-out items from the same public source benchmarks, with every RouterArena item excluded."
+  }
+}