feat: update to June 2026 models (Claude Sonnet 4.6, Gemini 3.5 Flash), add glossary button, and implement cost factor quota & vision fallback

routes/translate_routes.py

@@ -45,7 +45,7 @@ from config import config
 from translators import ExcelTranslator, WordTranslator, PowerPointTranslator
 from models.subscription import PlanType
 from middleware.tier_quota import tier_quota_service
-from services.auth_service import record_usage
+from services.auth_service import record_usage, check_usage_limits
 from middleware.validation import FileValidator, ValidationError, LanguageValidator, webhook_validator
 from middleware.api_key_auth import get_authenticated_user, get_user_from_api_key
 from utils import file_handler
@@ -135,9 +135,11 @@ file_handler_util = FileHandler()
 
 
 def _tier_for_quota(plan) -> str:
-    """Map plan to quota tier: pro (and equivalent) = unlimited, else free."""
+    """Map plan to quota tier: pro/starter (and equivalent) = unlimited, else free."""
     if plan in (PlanType.PRO, PlanType.BUSINESS, PlanType.ENTERPRISE):
         return "pro"
+    if plan == PlanType.STARTER:
+        return "starter"
     return "free"
 
 
@@ -630,6 +632,22 @@ async def translate_document_v1(
                     },
                     headers={"Retry-After": str(retry_after)},
                 )
+            
+            # Strict database plan limit check (Starter, Pro, Business, Enterprise)
+            usage = check_usage_limits(current_user)
+            if not usage["can_translate"]:
+                raise HTTPException(
+                    status_code=429,
+                    detail={
+                        "error": "QUOTA_EXCEEDED",
+                        "message": f"Monthly limit reached ({usage['docs_used']}/{usage['docs_limit']} documents). Upgrade your plan for more.",
+                        "details": {
+                            "current_usage": usage["docs_used"],
+                            "limit": usage["docs_limit"],
+                            "tier": tier,
+                        },
+                    },
+                )
             rate_limit_remaining = quota.remaining
         else:
             rate_limit_remaining = -1
@@ -993,7 +1011,9 @@ async def _run_translation_job(
             return (admin_val or "").strip() or os.getenv(env_var, default)
 
         api_key = _cfg(_admin_cfg.openrouter.api_key, "OPENROUTER_API_KEY")
-        model = _cfg(_admin_cfg.openrouter.model, "OPENROUTER_MODEL", "deepseek/deepseek-v3.2")
+        model = _cfg(_admin_cfg.openrouter.model, "OPENROUTER_MODEL", "google/gemini-3.5-flash")
+        if model in ("deepseek/deepseek-v3.2", "google/gemini-2.0-flash-001"):
+            model = "google/gemini-3.5-flash"
 
         # Story 3.10: Retrieve and format glossary terms for LLM prompt
         glossary_terms = None
@@ -1075,7 +1095,7 @@ async def _run_translation_job(
             )
         elif _p == "openrouter_premium":
             premium_key = _cfg(_admin_cfg.openrouter_premium.api_key, "OPENROUTER_API_KEY")
-            premium_model = _cfg(_admin_cfg.openrouter_premium.model, "OPENROUTER_PREMIUM_MODEL", "anthropic/claude-3.5-haiku")
+            premium_model = _cfg(_admin_cfg.openrouter_premium.model, "OPENROUTER_PREMIUM_MODEL", "anthropic/claude-sonnet-4.6")
             if not premium_key:
                 premium_key = api_key  # fall back to main openrouter key
             if premium_key:
@@ -1096,8 +1116,15 @@ async def _run_translation_job(
                 )
         elif _p == "deepseek":
             ds_key = _cfg(getattr(_admin_cfg, "deepseek", None) and _admin_cfg.deepseek.api_key, "DEEPSEEK_API_KEY")
-            ds_model = _cfg(getattr(_admin_cfg, "deepseek", None) and _admin_cfg.deepseek.model, "DEEPSEEK_MODEL", "deepseek-chat")
-            if ds_key:
+            ds_model = _cfg(getattr(_admin_cfg, "deepseek", None) and _admin_cfg.deepseek.model, "DEEPSEEK_MODEL", "deepseek/deepseek-chat")
+            if not ds_key and api_key:
+                translation_provider = OpenAITranslationProvider(
+                    api_key=api_key,
+                    model=ds_model,
+                    base_url="https://openrouter.ai/api/v1",
+                    timeout=int(os.getenv("OPENROUTER_TIMEOUT", "60")),
+                )
+            elif ds_key:
                 translation_provider = DeepSeekTranslationProvider(
                     api_key=ds_key,
                     model=ds_model,
@@ -1295,13 +1322,30 @@ async def _run_translation_job(
                 )
 
         if user_id:
-            await tier_quota_service.increment_on_success(user_id)
+            # Determine cost factor based on selected provider and model
+            cost_factor = 1
+            provider_lower = (provider or "").lower()
+            
+            prov_model = ""
+            if translation_provider:
+                prov_model = getattr(translation_provider, "model", "") or ""
+            
+            prov_model_lower = prov_model.lower()
+            if any(k in prov_model_lower for k in ["claude", "fable", "gpt-4"]) or provider_lower == "openrouter_premium":
+                if "haiku" in prov_model_lower:
+                    cost_factor = 1
+                else:
+                    cost_factor = 5
+
+            for _ in range(cost_factor):
+                await tier_quota_service.increment_on_success(user_id)
+
             # Persist monthly usage counters in PostgreSQL (docs + pages)
             pages = await asyncio.to_thread(
                 _estimate_pages, input_path, file_extension
             )
-            await asyncio.to_thread(record_usage, user_id, pages)
-            logger.info(f"Job {job_id}: usage recorded — {pages} page(s)")
+            await asyncio.to_thread(record_usage, user_id, pages, False, cost_factor)
+            logger.info(f"Job {job_id}: usage recorded — {pages} page(s) with cost factor {cost_factor}")
 
             # Apply watermark for Free-tier users
             plan_name = (user_plan or "free").lower()