diff --git a/models.yaml b/models.yaml index b9eb480..0054c4a 100644 --- a/models.yaml +++ b/models.yaml @@ -3,6 +3,22 @@ # Hand-owned behavior lives in quirks.yaml; this file is the embedded offline fallback. - provider: openai models: + - name: gpt-6.1-sol + type: chat + max_input_tokens: 1050000 + input_price: 2.0 + output_price: 10.0 + cache_read_price: 0.1 + cache_write_price: 2.5 + max_output_tokens: 128000 + supports_vision: true + supports_function_calling: true + reasoning_levels: + - low + - medium + - high + - xhigh + - max - name: gpt-6-luna type: chat max_input_tokens: 1050000 @@ -897,9 +913,9 @@ - name: gemma-4-26b-a4b-it type: chat max_input_tokens: 262144 - input_price: 0.09 - output_price: 0.3 - cache_read_price: 0.05 + input_price: 0.0765 + output_price: 0.255 + cache_read_price: 0.0425 max_output_tokens: 32768 supports_vision: true supports_function_calling: true @@ -3347,6 +3363,34 @@ - high - xhigh - max + - name: global.xai.grok-4.7 + type: chat + max_input_tokens: 500000 + input_price: 2.0 + output_price: 6.0 + cache_read_price: 0.5 + max_output_tokens: 500000 + supports_vision: true + supports_function_calling: true + reasoning_levels: + - low + - medium + - high + - xhigh + - name: us.xai.grok-4.7 + type: chat + max_input_tokens: 500000 + input_price: 2.2 + output_price: 6.6 + cache_read_price: 0.55 + max_output_tokens: 500000 + supports_vision: true + supports_function_calling: true + reasoning_levels: + - low + - medium + - high + - xhigh - name: global.openai.gpt-6-astra type: chat max_input_tokens: 1050000 @@ -7498,7 +7542,7 @@ supports_function_calling: true - name: deepseek/deepseek-v4-flash-0731 type: chat - max_input_tokens: 1310720 + max_input_tokens: 1048576 input_price: 0.018 output_price: 0.32 cache_read_price: 0.018 @@ -7524,10 +7568,10 @@ - name: deepseek/deepseek-v4-pro-0813 type: chat max_input_tokens: 1048576 - input_price: 0.48 - output_price: 4.2 - cache_read_price: 0.384 - max_output_tokens: 943718 + input_price: 0.66 + output_price: 1.98 + cache_read_price: 0.022 + max_output_tokens: 393216 supports_function_calling: true - name: deepseek/deepseek-v4.1-flash type: chat @@ -7919,9 +7963,9 @@ - name: google/gemma-4-26b-a4b-it type: chat max_input_tokens: 262144 - input_price: 0.09 - output_price: 0.3 - cache_read_price: 0.05 + input_price: 0.0765 + output_price: 0.255 + cache_read_price: 0.0425 max_output_tokens: 235929 supports_vision: true supports_function_calling: true @@ -8135,10 +8179,10 @@ - name: meta/muse-glimmer-30b type: chat max_input_tokens: 131072 - input_price: 0.3 - output_price: 1.2 + input_price: 0.35 + output_price: 1.5 cache_read_price: 0.04 - max_output_tokens: 16384 + max_output_tokens: 117964 supports_vision: true supports_function_calling: true - name: meta/muse-spark-1.1 @@ -9535,6 +9579,46 @@ max_output_tokens: 128000 supports_vision: true supports_function_calling: true + - name: openai/gpt-6.1-sol + type: chat + max_input_tokens: 1050000 + input_price: 2.0 + output_price: 10.0 + cache_read_price: 0.1 + cache_write_price: 2.5 + max_output_tokens: 128000 + supports_vision: true + supports_function_calling: true + - name: openai/gpt-6.1-sol-pro + type: chat + max_input_tokens: 1050000 + input_price: 2.0 + output_price: 10.0 + cache_read_price: 0.1 + cache_write_price: 2.5 + max_output_tokens: 128000 + supports_vision: true + supports_function_calling: true + - name: openai/gpt-6.1-sol-pro:batch + type: chat + max_input_tokens: 1050000 + input_price: 1.0 + output_price: 5.0 + cache_read_price: 0.05 + cache_write_price: 1.25 + max_output_tokens: 128000 + supports_vision: true + supports_function_calling: true + - name: openai/gpt-6.1-sol:batch + type: chat + max_input_tokens: 1050000 + input_price: 1.0 + output_price: 5.0 + cache_read_price: 0.05 + cache_write_price: 1.25 + max_output_tokens: 128000 + supports_vision: true + supports_function_calling: true - name: openai/gpt-audio type: chat max_input_tokens: 128000 @@ -9884,9 +9968,9 @@ - name: qwen/qwen3-30b-a3b type: chat max_input_tokens: 131072 - input_price: 0.12 - output_price: 0.5 - max_output_tokens: 16384 + input_price: 0.13 + output_price: 0.52 + max_output_tokens: 8192 supports_function_calling: true - name: qwen/qwen3-30b-a3b-instruct-2507 type: chat @@ -10194,10 +10278,10 @@ - name: qwen/qwen3.8-27b type: chat max_input_tokens: 1000000 - input_price: 0.0428 - output_price: 4.4 - cache_read_price: 0.0342 - max_output_tokens: 235929 + input_price: 0.42 + output_price: 3.0 + cache_read_price: 0.085 + max_output_tokens: 131072 supports_vision: true supports_function_calling: true - name: qwen/qwen3.8-27b:free @@ -10656,7 +10740,7 @@ max_input_tokens: 131072 input_price: 0.3 output_price: 0.9 - cache_read_price: 0.05 + cache_read_price: 0.055 max_output_tokens: 32768 supports_vision: true supports_function_calling: true @@ -10702,14 +10786,14 @@ - name: z-ai/glm-5.2 type: chat max_input_tokens: 1048576 - input_price: 0.1949 - output_price: 4.4 - cache_read_price: 0.1559 + input_price: 0.1449 + output_price: 3.99 + cache_read_price: 0.1159 max_output_tokens: 943718 supports_function_calling: true - name: z-ai/glm-5.3 type: chat - max_input_tokens: 1310720 + max_input_tokens: 1048576 input_price: 1.4 output_price: 4.4 cache_read_price: 0.26 @@ -10717,7 +10801,7 @@ supports_function_calling: true - name: z-ai/glm-5.3-flash type: chat - max_input_tokens: 1310720 + max_input_tokens: 1048576 input_price: 0.15 output_price: 0.5 cache_read_price: 0.03 @@ -10819,14 +10903,14 @@ - name: ~deepseek/deepseek-pro-latest type: chat max_input_tokens: 1048576 - input_price: 0.096 - output_price: 1.9584 - cache_read_price: 0.0749 - max_output_tokens: 943718 + input_price: 0.15 + output_price: 3.5 + cache_read_price: 0.15 + max_output_tokens: 393216 supports_function_calling: true - name: ~deepseek/deepseek-v4-flash-latest type: chat - max_input_tokens: 1310720 + max_input_tokens: 1048576 input_price: 0.012 output_price: 1.25 cache_read_price: 0.01 @@ -10856,7 +10940,7 @@ type: chat max_input_tokens: 1048576 input_price: 0.4 - output_price: 10.0 + output_price: 9.0 cache_read_price: 0.4 max_output_tokens: 943718 supports_vision: true @@ -10895,7 +10979,7 @@ max_input_tokens: 1050000 input_price: 2.0 output_price: 10.0 - cache_read_price: 0.2 + cache_read_price: 0.1 cache_write_price: 2.5 max_output_tokens: 128000 supports_vision: true @@ -10921,7 +11005,7 @@ supports_function_calling: true - name: ~z-ai/glm-flash-latest type: chat - max_input_tokens: 1310720 + max_input_tokens: 1048576 input_price: 0.02 output_price: 0.3 cache_read_price: 0.01 @@ -10930,10 +11014,10 @@ supports_function_calling: true - name: ~z-ai/glm-latest type: chat - max_input_tokens: 1310720 - input_price: 0.19 + max_input_tokens: 1048576 + input_price: 0.12 output_price: 4.0 - cache_read_price: 0.19 + cache_read_price: 0.12 max_output_tokens: 131072 supports_function_calling: true - provider: github @@ -11042,7 +11126,7 @@ - name: XiaomiMiMo/MiMo-V2.6-Flash type: chat max_input_tokens: 1048576 - input_price: 0.14 + input_price: 0.08 output_price: 0.28 cache_read_price: 0.0028 max_output_tokens: 131072