feat(together_ai): capability baseline, reasoning controls, and a current model map

README.md

@@ -360,7 +360,7 @@ curl -X POST 'http://0.0.0.0:4000/v1/chat/completions' \
 | [Snowflake (`snowflake`)](https://docs.litellm.ai/docs/providers/snowflake) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
 | [Text Completion Codestral (`text-completion-codestral`)](https://docs.litellm.ai/docs/providers/codestral) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
 | [Text Completion OpenAI (`text-completion-openai`)](https://docs.litellm.ai/docs/providers/text_completion_openai) | ✅ | ✅ | ✅ |  |  | ✅ | ✅ | ✅ | ✅ |  |
-| [Together AI (`together_ai`)](https://docs.litellm.ai/docs/providers/togetherai) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
+| [Together AI (`together_ai`)](https://docs.litellm.ai/docs/providers/togetherai) | ✅ | ✅ | ✅ | ✅ |  |  |  |  |  |  |
 | [Topaz (`topaz`)](https://docs.litellm.ai/docs/providers/topaz) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
 | [Triton (`triton`)](https://docs.litellm.ai/docs/providers/triton-inference-server) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
 | [V0 (`v0`)](https://docs.litellm.ai/docs/providers/v0) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |

litellm/constants.py

@@ -746,6 +746,7 @@
     "api.groq.com/openai/v1",
     "https://integrate.api.nvidia.com/v1",
     "api.deepseek.com/v1",
+    "api.together.ai/v1",
     "api.together.xyz/v1",
     "app.empower.dev/api/v1",
     "https://api.friendli.ai/serverless/v1",

litellm/litellm_core_utils/get_llm_provider_logic.py

@@ -704,13 +704,10 @@ def _get_openai_compatible_provider_info(
             dynamic_api_key,
         ) = litellm.ZAIChatConfig()._get_openai_compatible_provider_info(api_base, api_key)
     elif custom_llm_provider == "together_ai":
-        api_base = api_base or get_secret_str("TOGETHER_AI_API_BASE") or "https://api.together.xyz/v1"
-        dynamic_api_key = api_key or (
-            get_secret_str("TOGETHER_API_KEY")
-            or get_secret_str("TOGETHER_AI_API_KEY")
-            or get_secret_str("TOGETHERAI_API_KEY")
-            or get_secret_str("TOGETHER_AI_TOKEN")
-        )
+        (
+            api_base,
+            dynamic_api_key,
+        ) = litellm.TogetherAIConfig()._get_openai_compatible_provider_info(api_base=api_base, api_key=api_key)
     elif custom_llm_provider == "friendliai":
         api_base = api_base or get_secret("FRIENDLI_API_BASE") or "https://api.friendli.ai/serverless/v1"
         dynamic_api_key = api_key or get_secret_str("FRIENDLIAI_API_KEY") or get_secret_str("FRIENDLI_TOKEN")

litellm/litellm_core_utils/llm_cost_calc/utils.py

@@ -309,7 +309,11 @@ def _get_token_base_cost(
         float,
         _get_cost_per_unit(model_info, "cache_creation_input_token_cost_above_1hr"),
     )
-    cache_read_cost = cast(float, _get_cost_per_unit(model_info, cache_read_cost_key))
+    # A model that publishes no cached-input rate gives no cached discount, so a cache hit
+    # bills at the standard input rate; read as 0.0 it would bill the hit as free. This is
+    # what the tiered-pricing path above already does for cache reads.
+    published_cache_read_cost: Final = _get_cost_per_unit(model_info, cache_read_cost_key, None)
+    cache_read_cost = prompt_base_cost if published_cache_read_cost is None else cast(float, published_cache_read_cost)
 
     ## CHECK IF ABOVE THRESHOLD
     # Optimization: collect threshold keys first to avoid sorting all model_info keys.
@@ -937,19 +941,25 @@ def generic_cost_per_token(
     # 1. If text_tokens is explicitly provided and > 0, use it
     # 2. If there's a breakdown (reasoning/audio/image/video tokens), calculate text_tokens as the remainder
     # 3. If no breakdown at all, assume all completion_tokens are text_tokens
+    # 4. If text_tokens is the total with reasoning nested inside it, take the remainder too:
+    #    Together AI reports that shape on Qwen3.6/3.7 Plus (text_tokens == completion_tokens
+    #    alongside a non-zero reasoning_tokens) and the disjoint shape on Qwen3.7 Max, and
+    #    reasoning is billed on its own below, so the overlap charged those tokens twice.
+    #    Mirrors the prompt-side guard above; scoped to reasoning, the overlap seen on the wire.
     has_token_breakdown: Final = image_tokens > 0 or audio_tokens > 0 or reasoning_tokens > 0 or video_tokens > 0
-    if text_tokens == 0:
-        if has_token_breakdown:
-            # Calculate text tokens as remainder when we have a breakdown
-            # This handles cases like OpenAI's reasoning models where text_tokens isn't provided
-            text_tokens = max(
-                0,
-                usage.completion_tokens - reasoning_tokens - audio_tokens - image_tokens - video_tokens,
-            )
-        else:
-            # No breakdown at all, all tokens are text tokens
-            text_tokens = usage.completion_tokens
-            is_text_tokens_total = True
+    reasoning_overlaps_text: Final = reasoning_tokens > 0 and text_tokens + reasoning_tokens > usage.completion_tokens
+    if text_tokens == 0 and not has_token_breakdown:
+        # No breakdown at all, all tokens are text tokens
+        text_tokens = usage.completion_tokens
+        is_text_tokens_total = True
+    elif text_tokens == 0 or reasoning_overlaps_text:
+        # Calculate text tokens as remainder when we have a breakdown
+        # This handles cases like OpenAI's reasoning models where text_tokens isn't provided
+        text_tokens = max(
+            0,
+            usage.completion_tokens - reasoning_tokens - audio_tokens - image_tokens - video_tokens,
+        )
+
     ## TEXT COST
     completion_cost = float(text_tokens) * completion_base_cost
 

litellm/litellm_core_utils/prompt_templates/factory.py

@@ -649,7 +649,7 @@ def get_model_info(token, model):
     try:
         headers: Final = {"Authorization": f"Bearer {token}"}
         client: Final = HTTPHandler(concurrent_limit=1)
-        response: Final = client.get("https://api.together.xyz/models/info", headers=headers)
+        response: Final = client.get("https://api.together.ai/models/info", headers=headers)
         if response.status_code == 200:
             model_info: Final = response.json()
             for m in model_info: