From e148868f0c773ef933bba9d84cc3e2d34c572554 Mon Sep 17 00:00:00 2001 From: mateo Date: Wed, 2 Sep 2026 20:38:45 +0000 Subject: [PATCH] fix(model_prices): set watsonx max_output_tokens from IBM's documented maximum new tokens Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- ...odel_prices_and_context_window_backup.json | 24 +++++++++---------- model_prices_and_context_window.json | 24 +++++++++---------- 2 files changed, 24 insertions(+), 24 deletions(-) diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 2e717b52b28..1935bc97c3b 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -46960,9 +46960,9 @@ "supports_vision": false }, "watsonx/ibm/granite-4-h-small": { - "max_tokens": 20480, + "max_tokens": 131072, "max_input_tokens": 131072, - "max_output_tokens": 20480, + "max_output_tokens": 131072, "input_cost_per_token": 6.36e-08, "output_cost_per_token": 2.65e-07, "litellm_provider": "watsonx", @@ -47093,9 +47093,9 @@ "supports_vision": true }, "watsonx/meta-llama/llama-3-3-70b-instruct": { - "max_tokens": 128000, + "max_tokens": 131072, "max_input_tokens": 131072, - "max_output_tokens": 128000, + "max_output_tokens": 131072, "input_cost_per_token": 7.526e-07, "output_cost_per_token": 7.526e-07, "litellm_provider": "watsonx", @@ -47106,9 +47106,9 @@ "source": "https://dataplatform.cloud.ibm.com/docs/content/wsj/analyze-data/fm-models.html?context=wx" }, "watsonx/meta-llama/llama-4-maverick-17b": { - "max_tokens": 128000, + "max_tokens": 8192, "max_input_tokens": 131072, - "max_output_tokens": 128000, + "max_output_tokens": 8192, "input_cost_per_token": 3.71e-07, "output_cost_per_token": 1.484e-06, "litellm_provider": "watsonx", @@ -47119,9 +47119,9 @@ "source": "https://dataplatform.cloud.ibm.com/docs/content/wsj/analyze-data/fm-models.html?context=wx" }, "watsonx/meta-llama/llama-4-maverick-17b-128e-instruct-fp8": { - "max_tokens": 128000, + "max_tokens": 8192, "max_input_tokens": 131072, - "max_output_tokens": 128000, + "max_output_tokens": 8192, "input_cost_per_token": 3.71e-07, "output_cost_per_token": 1.484e-06, "litellm_provider": "watsonx", @@ -47168,9 +47168,9 @@ "supports_vision": false }, "watsonx/mistralai/mistral-small-3-1-24b-instruct-2503": { - "max_tokens": 32000, + "max_tokens": 16384, "max_input_tokens": 131072, - "max_output_tokens": 32000, + "max_output_tokens": 16384, "input_cost_per_token": 1.06e-07, "output_cost_per_token": 3.18e-07, "litellm_provider": "watsonx", @@ -47193,9 +47193,9 @@ "supports_vision": true }, "watsonx/openai/gpt-oss-120b": { - "max_tokens": 8192, + "max_tokens": 131072, "max_input_tokens": 131072, - "max_output_tokens": 8192, + "max_output_tokens": 131072, "input_cost_per_token": 1.59e-07, "output_cost_per_token": 6.36e-07, "litellm_provider": "watsonx", diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 2e717b52b28..1935bc97c3b 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -46960,9 +46960,9 @@ "supports_vision": false }, "watsonx/ibm/granite-4-h-small": { - "max_tokens": 20480, + "max_tokens": 131072, "max_input_tokens": 131072, - "max_output_tokens": 20480, + "max_output_tokens": 131072, "input_cost_per_token": 6.36e-08, "output_cost_per_token": 2.65e-07, "litellm_provider": "watsonx", @@ -47093,9 +47093,9 @@ "supports_vision": true }, "watsonx/meta-llama/llama-3-3-70b-instruct": { - "max_tokens": 128000, + "max_tokens": 131072, "max_input_tokens": 131072, - "max_output_tokens": 128000, + "max_output_tokens": 131072, "input_cost_per_token": 7.526e-07, "output_cost_per_token": 7.526e-07, "litellm_provider": "watsonx", @@ -47106,9 +47106,9 @@ "source": "https://dataplatform.cloud.ibm.com/docs/content/wsj/analyze-data/fm-models.html?context=wx" }, "watsonx/meta-llama/llama-4-maverick-17b": { - "max_tokens": 128000, + "max_tokens": 8192, "max_input_tokens": 131072, - "max_output_tokens": 128000, + "max_output_tokens": 8192, "input_cost_per_token": 3.71e-07, "output_cost_per_token": 1.484e-06, "litellm_provider": "watsonx", @@ -47119,9 +47119,9 @@ "source": "https://dataplatform.cloud.ibm.com/docs/content/wsj/analyze-data/fm-models.html?context=wx" }, "watsonx/meta-llama/llama-4-maverick-17b-128e-instruct-fp8": { - "max_tokens": 128000, + "max_tokens": 8192, "max_input_tokens": 131072, - "max_output_tokens": 128000, + "max_output_tokens": 8192, "input_cost_per_token": 3.71e-07, "output_cost_per_token": 1.484e-06, "litellm_provider": "watsonx", @@ -47168,9 +47168,9 @@ "supports_vision": false }, "watsonx/mistralai/mistral-small-3-1-24b-instruct-2503": { - "max_tokens": 32000, + "max_tokens": 16384, "max_input_tokens": 131072, - "max_output_tokens": 32000, + "max_output_tokens": 16384, "input_cost_per_token": 1.06e-07, "output_cost_per_token": 3.18e-07, "litellm_provider": "watsonx", @@ -47193,9 +47193,9 @@ "supports_vision": true }, "watsonx/openai/gpt-oss-120b": { - "max_tokens": 8192, + "max_tokens": 131072, "max_input_tokens": 131072, - "max_output_tokens": 8192, + "max_output_tokens": 131072, "input_cost_per_token": 1.59e-07, "output_cost_per_token": 6.36e-07, "litellm_provider": "watsonx",