diff --git a/src/rotator_library/provider_config.py b/src/rotator_library/provider_config.py
index 51d40043..94630ca3 100644
--- a/src/rotator_library/provider_config.py
+++ b/src/rotator_library/provider_config.py
@@ -677,6 +677,16 @@ def _load_api_bases(self) -> None:
                         f"Detected API base override for {provider}: {value}"
                     )
 
+        # Handle firmware provider: has default api_base but LiteLLM doesn't
+        # recognize it as a native provider. Without this, FIRMWARE_API_KEY_*
+        # users without explicit FIRMWARE_API_BASE would have requests fail.
+        if "firmware" not in self._api_bases and "firmware" not in KNOWN_PROVIDERS:
+            self._api_bases["firmware"] = "https://app.firmware.ai/api/v1"
+            self._custom_providers.add("firmware")
+            lib_logger.info(
+                "Using default api_base for firmware: https://app.firmware.ai/api/v1"
+            )
+
     def is_known_provider(self, provider: str) -> bool:
         """Check if provider is known to LiteLLM."""
         return provider.lower() in KNOWN_PROVIDERS
diff --git a/src/rotator_library/providers/firmware_provider.py b/src/rotator_library/providers/firmware_provider.py
index e71316fa..978ae3e0 100644
--- a/src/rotator_library/providers/firmware_provider.py
+++ b/src/rotator_library/providers/firmware_provider.py
@@ -32,8 +32,18 @@
 class FirmwareProvider(FirmwareQuotaTracker, ProviderInterface):
     """
     Provider implementation for the Firmware.ai API with quota tracking.
+
+    Firmware.ai is OpenAI-compatible, so requests are routed through LiteLLM's
+    OpenAI provider with api_base override. This class provides:
+    - Quota tracking via the FirmwareQuotaTracker mixin
+    - Model discovery from Firmware.ai's /models endpoint
+    - Cost calculation is skipped since Firmware models aren't in LiteLLM's pricing DB
     """
 
+    # Skip LiteLLM cost calculation - Firmware.ai models use custom naming
+    # (e.g., firmware/anthropic/claude-sonnet-4-5) not in LiteLLM's pricing database
+    skip_cost_calculation: bool = True
+
     # Quota groups for tracking 5-hour rolling window limits
     # Uses a virtual model "firmware/_quota" for credential-level quota tracking
     model_quota_groups = {