diff --git a/litellm/llms/fireworks_ai/chat/transformation.py b/litellm/llms/fireworks_ai/chat/transformation.py
index f79e9467041..2081c80a987 100644
--- a/litellm/llms/fireworks_ai/chat/transformation.py
+++ b/litellm/llms/fireworks_ai/chat/transformation.py
@@ -261,9 +261,7 @@ class FireworksAIConfig(OpenAIGPTConfig):
         return {
             "supports_function_calling": True,
             "supports_tool_choice": True,
-            "supports_prompt_caching": True,  # https://docs.fireworks.ai/guides/prompt-caching
-            "supports_pdf_input": True,  # via document inlining
-            "supports_vision": True,  # via document inlining
+            "supports_pdf_input": True,
         }
 
     def transform_request(
@@ -413,7 +411,14 @@ class FireworksAIConfig(OpenAIGPTConfig):
                 "(or FIREWORKS_AI_API_KEY / FIREWORKSAI_API_KEY / FIREWORKS_AI_TOKEN)."
             )
 
-        base_url = "https://api.fireworks.ai"
+        base_url = (
+            api_base
+            or get_secret_str("FIREWORKS_API_BASE")
+            or "https://api.fireworks.ai"
+        )
+        if base_url.endswith("/inference/v1"):
+            base_url = base_url[: -len("/inference/v1")]
+        base_url = base_url.rstrip("/")
         headers = {"Authorization": f"Bearer {api_key}"}
         seen: set = set()
         result: List[str] = []
diff --git a/litellm/llms/fireworks_ai/cost_calculator.py b/litellm/llms/fireworks_ai/cost_calculator.py
index 27152f8936b..8e64d902010 100644
--- a/litellm/llms/fireworks_ai/cost_calculator.py
+++ b/litellm/llms/fireworks_ai/cost_calculator.py
@@ -4,6 +4,7 @@ For calculating cost of fireworks ai serverless inference models.
 
 from typing import Tuple
 
+from litellm._logging import verbose_logger
 from litellm.constants import (
     FIREWORKS_AI_4_B,
     FIREWORKS_AI_16_B,
@@ -74,7 +75,13 @@ def cost_per_token(model: str, usage: Usage) -> Tuple[float, float]:
         return generic_cost_per_token(
             model=model, usage=usage, custom_llm_provider="fireworks_ai"
         )
-    except Exception:
+    except Exception as e:
+        verbose_logger.debug(
+            "fireworks_ai cost_per_token: model '%s' not in pricing JSON, "
+            "falling back to size heuristic: %s",
+            model,
+            e,
+        )
         base_model = get_base_model_for_pricing(model_name=model)
         return generic_cost_per_token(
             model=base_model, usage=usage, custom_llm_provider="fireworks_ai"
diff --git a/tests/test_litellm/llms/fireworks_ai/chat/test_fireworks_ai_chat_transformation.py b/tests/test_litellm/llms/fireworks_ai/chat/test_fireworks_ai_chat_transformation.py
index d13c594a024..591c5c82fbd 100644
--- a/tests/test_litellm/llms/fireworks_ai/chat/test_fireworks_ai_chat_transformation.py
+++ b/tests/test_litellm/llms/fireworks_ai/chat/test_fireworks_ai_chat_transformation.py
@@ -94,17 +94,23 @@ def test_supports_reasoning_effort():
 
 
 def test_get_supported_openai_params_reasoning_effort():
-    """Test that reasoning_effort is always included — Fireworks accepts it on all models."""
+    """
+    Test that reasoning_effort is always included in supported params.
+
+    Verified against live Fireworks API (Apr 2026): sending reasoning_effort
+    to non-reasoning models (e.g. llama-v3p3-70b-instruct) returns a
+    successful response — the API accepts and silently ignores the parameter.
+    Models that do support it (qwen3, deepseek-v3p1/v3p2, glm-5p1) use it
+    to control reasoning depth. The Fireworks API never returns a 4xx for an
+    unrecognised reasoning_effort value on unsupported models.
+    """
     config = FireworksAIConfig()
 
-    # reasoning_effort should be present for reasoning models
     supported_params = config.get_supported_openai_params(
         "fireworks_ai/accounts/fireworks/models/qwen3-8b"
     )
     assert "reasoning_effort" in supported_params
 
-    # reasoning_effort should also be present for non-reasoning models
-    # (Fireworks API accepts it; unsupported models simply ignore it)
     other_params = config.get_supported_openai_params(
         "fireworks_ai/accounts/fireworks/models/llama-v3-70b-instruct"
     )
@@ -349,6 +355,80 @@ def test_get_models_account_id_equals_fireworks_no_duplicate_query():
         assert result == ["fireworks_ai/accounts/fireworks/models/deepseek-v3p2"]
 
 
+@pytest.mark.parametrize(
+    "api_base, expected_base",
+    [
+        (
+            "https://my-proxy.example.com",
+            "https://my-proxy.example.com/v1/accounts/fireworks/models",
+        ),
+        (
+            "https://my-proxy.example.com/inference/v1",
+            "https://my-proxy.example.com/v1/accounts/fireworks/models",
+        ),
+        (
+            "https://my-proxy.example.com/custom/",
+            "https://my-proxy.example.com/custom/v1/accounts/fireworks/models",
+        ),
+        (None, "https://api.fireworks.ai/v1/accounts/fireworks/models"),
+    ],
+    ids=["plain-base", "inference-v1-stripped", "trailing-slash", "default"],
+)
+def test_get_models_respects_api_base(api_base, expected_base):
+    """get_models should use api_base (or FIREWORKS_API_BASE) instead of hardcoding."""
+    config = FireworksAIConfig()
+
+    mock_response = MagicMock()
+    mock_response.status_code = 200
+    mock_response.json.return_value = {
+        "models": [{"name": "accounts/fireworks/models/llama-v3-70b"}]
+    }
+
+    with (
+        patch(
+            "litellm.module_level_client.get", return_value=mock_response
+        ) as mock_get,
+        patch(
+            "litellm.llms.fireworks_ai.chat.transformation.get_secret_str",
+            return_value=None,
+        ),
+    ):
+        config.get_models(api_key="test-key", api_base=api_base)
+
+        called_url = mock_get.call_args.kwargs.get("url") or mock_get.call_args[1].get(
+            "url", ""
+        )
+        assert called_url == expected_base
+        assert "/v1/v1/" not in called_url
+
+
+def test_get_models_respects_fireworks_api_base_env():
+    """get_models should fall back to FIREWORKS_API_BASE env when no api_base arg."""
+    config = FireworksAIConfig()
+
+    mock_response = MagicMock()
+    mock_response.status_code = 200
+    mock_response.json.return_value = {"models": []}
+
+    with (
+        patch(
+            "litellm.module_level_client.get", return_value=mock_response
+        ) as mock_get,
+        patch(
+            "litellm.llms.fireworks_ai.chat.transformation.get_secret_str",
+            side_effect=lambda key: (
+                "https://env-proxy.example.com" if key == "FIREWORKS_API_BASE" else None
+            ),
+        ),
+    ):
+        config.get_models(api_key="test-key")
+
+        called_url = mock_get.call_args.kwargs.get("url") or mock_get.call_args[1].get(
+            "url", ""
+        )
+        assert called_url.startswith("https://env-proxy.example.com/")
+
+
 def test_transform_messages_helper_removes_provider_specific_fields():
     """
     Test that _transform_messages_helper removes provider_specific_fields from messages.
diff --git a/tests/test_litellm/llms/fireworks_ai/responses/test_fireworks_ai_responses_transformation.py b/tests/test_litellm/llms/fireworks_ai/responses/test_fireworks_ai_responses_transformation.py
index 7b8c4e11c2e..be2af7d8d3c 100644
--- a/tests/test_litellm/llms/fireworks_ai/responses/test_fireworks_ai_responses_transformation.py
+++ b/tests/test_litellm/llms/fireworks_ai/responses/test_fireworks_ai_responses_transformation.py
@@ -85,6 +85,15 @@ class TestValidateEnvironment:
             )
             assert headers["Authorization"] == "Bearer env-key"
 
+    def test_should_not_overwrite_existing_authorization(self, config):
+        params = GenericLiteLLMParams(api_key="new-key")
+        headers = config.validate_environment(
+            headers={"Authorization": "Bearer pre-existing"},
+            model="accounts/fireworks/models/llama-v3-70b",
+            litellm_params=params,
+        )
+        assert headers["Authorization"] == "Bearer pre-existing"
+
 
 class TestGetCompleteUrl:
     """Tests for get_complete_url."""