diff --git a/models.yaml b/models.yaml index efafb1b..1744b99 100644 --- a/models.yaml +++ b/models.yaml @@ -1815,6 +1815,20 @@ use_pkce_in_device_flow: false include_state_in_token_exchange: true models: + - name: 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: grok-4.6 type: chat max_input_tokens: 500000 @@ -5050,6 +5064,10 @@ require_max_tokens: true supports_vision: true supports_function_calling: true + reasoning_levels: + - low + - high + - max - name: '@cf/qwen/qwen3.8-27b' type: chat max_input_tokens: 262144 @@ -5070,13 +5088,13 @@ input_price: 1.4 output_price: 4.4 cache_read_price: 0.26 - max_output_tokens: 1310720 + max_output_tokens: 1048576 require_max_tokens: true supports_function_calling: true reasoning_levels: - low - - medium - high + - max - name: '@cf/deepseek-ai/deepseek-v4-pro-0813' type: chat max_input_tokens: 1048576 @@ -5087,6 +5105,8 @@ require_max_tokens: true supports_function_calling: true reasoning_levels: + - none + - low - high - max - name: '@cf/deepseek-ai/deepseek-v4-flash-0731' @@ -5099,6 +5119,8 @@ require_max_tokens: true supports_function_calling: true reasoning_levels: + - none + - low - high - max - name: '@cf/zai-org/glm-5.2' @@ -5111,9 +5133,9 @@ require_max_tokens: true supports_function_calling: true reasoning_levels: - - low - - medium + - none - high + - max - name: '@cf/moonshotai/kimi-k2.7-code' type: chat max_input_tokens: 262144 @@ -5124,10 +5146,6 @@ require_max_tokens: true supports_vision: true supports_function_calling: true - reasoning_levels: - - low - - medium - - high - name: '@cf/moonshotai/kimi-k2.6' type: chat max_input_tokens: 262144 @@ -5139,8 +5157,7 @@ supports_vision: true supports_function_calling: true reasoning_levels: - - low - - medium + - none - high - name: '@cf/google/gemma-4-26b-a4b-it' type: chat @@ -5152,8 +5169,7 @@ supports_vision: true supports_function_calling: true reasoning_levels: - - low - - medium + - none - high - name: '@cf/nvidia/nemotron-3-120b-a12b' type: chat @@ -5175,10 +5191,6 @@ max_output_tokens: 131072 require_max_tokens: true supports_function_calling: true - reasoning_levels: - - low - - medium - - high - name: '@cf/ibm-granite/granite-4.0-h-micro' type: chat max_input_tokens: 131000 @@ -6002,6 +6014,7 @@ max_output_tokens: 384000 supports_function_calling: true reasoning_levels: + - low - high - max - name: deepseek-chat @@ -6163,6 +6176,16 @@ max_output_tokens: 32768 supports_vision: true supports_function_calling: true + - name: glm-4.6v-flash + type: chat + max_input_tokens: 65536 + input_price: 0.0 + output_price: 0.0 + cache_read_price: 0.0 + cache_write_price: 0.0 + max_output_tokens: 32768 + supports_vision: true + supports_function_calling: true - name: glm-4.6 type: chat max_input_tokens: 204800 @@ -6217,12 +6240,6 @@ type: reranker max_input_tokens: 4096 input_price: 0.112 - - name: glm-4.6v-flash - type: chat - max_input_tokens: 65536 - input_price: 0.0 - output_price: 0.0 - supports_vision: true - provider: minimax models: - name: MiniMax-M3 @@ -6600,17 +6617,6 @@ require_max_tokens: true supports_vision: true supports_function_calling: true - - name: anthropic/claude-opus-4 - type: chat - max_input_tokens: 200000 - input_price: 15.0 - output_price: 75.0 - cache_read_price: 1.5 - cache_write_price: 18.75 - max_output_tokens: 32000 - require_max_tokens: true - supports_vision: true - supports_function_calling: true - name: anthropic/claude-opus-4.1 type: chat max_input_tokens: 200000 @@ -6758,7 +6764,7 @@ supports_function_calling: true - name: anthropic/claude-sonnet-4 type: chat - max_input_tokens: 1000000 + max_input_tokens: 200000 input_price: 3.0 output_price: 15.0 cache_read_price: 0.3 @@ -7034,7 +7040,7 @@ type: chat max_input_tokens: 1310720 input_price: 0.04 - output_price: 0.16 + output_price: 0.64 cache_read_price: 0.016 max_output_tokens: 943718 supports_function_calling: true @@ -7520,9 +7526,9 @@ - name: ibm-granite/granite-4.2-8b type: chat max_input_tokens: 131072 - input_price: 0.1 - output_price: 0.15 - cache_read_price: 0.05 + input_price: 0.06 + output_price: 0.25 + cache_read_price: 0.015 max_output_tokens: 117964 supports_function_calling: true - name: inception/mercury-2 @@ -7642,9 +7648,9 @@ - name: meta-llama/llama-3.1-70b-instruct type: chat max_input_tokens: 131072 - input_price: 0.72 - output_price: 0.72 - max_output_tokens: 8192 + input_price: 0.4 + output_price: 0.4 + max_output_tokens: 16384 supports_function_calling: true - name: meta-llama/llama-3.1-8b-instruct type: chat @@ -7676,8 +7682,8 @@ - name: meta-llama/llama-4-maverick type: chat max_input_tokens: 1048576 - input_price: 0.2 - output_price: 0.8 + input_price: 0.1875 + output_price: 0.6525 max_output_tokens: 16384 supports_vision: true supports_function_calling: true @@ -7832,15 +7838,6 @@ max_output_tokens: 512000 supports_vision: true supports_function_calling: true - - name: minimax/minimax-m3:batch - type: chat - max_input_tokens: 524288 - input_price: 0.3 - output_price: 1.2 - cache_read_price: 0.06 - max_output_tokens: 471859 - supports_vision: true - supports_function_calling: true - name: mistralai/codestral-2508 type: chat max_input_tokens: 256000 @@ -8090,15 +8087,6 @@ supports_vision: true supports_function_calling: true - name: moonshotai/kimi-k3 - type: chat - max_input_tokens: 1048576 - input_price: 1.7 - output_price: 8.5 - cache_read_price: 0.17 - max_output_tokens: 943718 - supports_vision: true - supports_function_calling: true - - name: moonshotai/kimi-k3:batch type: chat max_input_tokens: 1048576 input_price: 3.0 @@ -8156,8 +8144,9 @@ - name: nvidia/nemotron-3-nano-30b-a3b type: chat max_input_tokens: 262144 - input_price: 0.06 - output_price: 0.24 + input_price: 0.05 + output_price: 0.2 + cache_read_price: 0.03 max_output_tokens: 235929 supports_function_calling: true - name: nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free @@ -9057,13 +9046,6 @@ cache_read_price: 0.075 max_output_tokens: 65536 supports_function_calling: true - - name: openai/gpt-oss-120b:batch - type: chat - max_input_tokens: 131072 - input_price: 0.15 - output_price: 0.6 - max_output_tokens: 117964 - supports_function_calling: true - name: openai/gpt-oss-20b type: chat max_input_tokens: 131072 @@ -9562,14 +9544,6 @@ max_output_tokens: 32768 supports_vision: true supports_function_calling: true - - name: qwen/qwen3.5-9b:batch - type: chat - max_input_tokens: 262144 - input_price: 0.17 - output_price: 0.25 - max_output_tokens: 235929 - supports_vision: true - supports_function_calling: true - name: qwen/qwen3.5-flash-02-23 type: chat max_input_tokens: 1000000 @@ -9676,20 +9650,12 @@ cache_read_price: 0.25 max_output_tokens: 131072 supports_function_calling: true - - name: qwen/qwen3.8-2.4t-a95b:batch - type: chat - max_input_tokens: 1010000 - input_price: 2.0 - output_price: 6.0 - cache_read_price: 0.25 - max_output_tokens: 909000 - supports_function_calling: true - name: qwen/qwen3.8-27b type: chat max_input_tokens: 1000000 - input_price: 0.2 - output_price: 2.5 - cache_read_price: 0.05 + input_price: 0.42 + output_price: 3.0 + cache_read_price: 0.085 max_output_tokens: 131072 supports_vision: true supports_function_calling: true @@ -9913,15 +9879,6 @@ max_output_tokens: 262144 supports_vision: true supports_function_calling: true - - name: thinkingmachines/inkling:batch - type: chat - max_input_tokens: 524288 - input_price: 1.0 - output_price: 4.05 - cache_read_price: 0.17 - max_output_tokens: 471859 - supports_vision: true - supports_function_calling: true - name: thinkingmachines/inkling:free type: chat max_input_tokens: 1048576 @@ -10020,6 +9977,15 @@ max_output_tokens: 450000 supports_vision: true supports_function_calling: true + - name: x-ai/grok-4.7 + type: chat + max_input_tokens: 500000 + input_price: 1.6 + output_price: 4.8 + cache_read_price: 0.4 + max_output_tokens: 450000 + supports_vision: true + supports_function_calling: true - name: x-ai/grok-build-0.1 type: chat max_input_tokens: 256000 @@ -10046,6 +10012,33 @@ cache_read_price: 0.0036 max_output_tokens: 131072 supports_function_calling: true + - name: xiaomi/mimo-v2.6-flash + type: chat + max_input_tokens: 1048576 + input_price: 0.14 + output_price: 0.28 + cache_read_price: 0.0028 + max_output_tokens: 131072 + supports_vision: true + supports_function_calling: true + - name: xiaomi/mimo-v2.6-pro + type: chat + max_input_tokens: 1048576 + input_price: 0.435 + output_price: 0.87 + cache_read_price: 0.0036 + max_output_tokens: 131072 + supports_vision: true + supports_function_calling: true + - name: xiaomi/mimo-v2.6-pro-ultraspeed + type: chat + max_input_tokens: 1048576 + input_price: 4.35 + output_price: 8.7 + cache_read_price: 0.036 + max_output_tokens: 131072 + supports_vision: true + supports_function_calling: true - name: z-ai/glm-4.5 type: chat max_input_tokens: 131072 @@ -10152,18 +10145,18 @@ - name: z-ai/glm-5.3 type: chat max_input_tokens: 1310720 - input_price: 0.91 - output_price: 2.86 - cache_read_price: 0.169 + input_price: 0.84 + output_price: 2.64 + cache_read_price: 0.156 max_output_tokens: 131072 supports_function_calling: true - name: z-ai/glm-5.3-flash type: chat max_input_tokens: 1310720 - input_price: 0.09 - output_price: 0.3 - cache_read_price: 0.018 - max_output_tokens: 131072 + input_price: 0.15 + output_price: 0.5 + cache_read_price: 0.05 + max_output_tokens: 943718 supports_vision: true supports_function_calling: true - name: z-ai/glm-5.3-flash:batch @@ -10253,17 +10246,17 @@ - name: ~deepseek/deepseek-pro-latest type: chat max_input_tokens: 1048576 - input_price: 0.66 - output_price: 1.98 - cache_read_price: 0.022 + input_price: 0.5586 + output_price: 1.6759 + cache_read_price: 0.0186 max_output_tokens: 384000 supports_function_calling: true - name: ~deepseek/deepseek-v4-flash-latest type: chat max_input_tokens: 1310720 - input_price: 0.04 - output_price: 0.16 - cache_read_price: 0.016 + input_price: 0.03 + output_price: 1.0 + cache_read_price: 0.012 max_output_tokens: 943718 supports_function_calling: true - name: ~google/gemini-flash-latest @@ -10289,9 +10282,9 @@ - name: ~moonshotai/kimi-latest type: chat max_input_tokens: 1048576 - input_price: 1.7 - output_price: 8.5 - cache_read_price: 0.17 + input_price: 1.5 + output_price: 7.5 + cache_read_price: 0.15 max_output_tokens: 943718 supports_vision: true supports_function_calling: true @@ -10347,9 +10340,9 @@ - name: ~x-ai/grok-latest type: chat max_input_tokens: 500000 - input_price: 2.0 - output_price: 6.0 - cache_read_price: 0.5 + input_price: 1.6 + output_price: 4.8 + cache_read_price: 0.4 max_output_tokens: 450000 supports_vision: true supports_function_calling: true @@ -10358,17 +10351,17 @@ max_input_tokens: 1310720 input_price: 0.075 output_price: 0.25 - cache_read_price: 0.015 - max_output_tokens: 943718 + cache_read_price: 0.02 + max_output_tokens: 102400 supports_vision: true supports_function_calling: true - name: ~z-ai/glm-latest type: chat max_input_tokens: 1310720 - input_price: 0.7735 - output_price: 2.431 - cache_read_price: 0.1271 - max_output_tokens: 943718 + input_price: 0.7839 + output_price: 2.6532 + cache_read_price: 0.1568 + max_output_tokens: 235929 supports_function_calling: true - provider: github models: @@ -10623,9 +10616,9 @@ - name: tencent/Hy3 type: chat max_input_tokens: 262144 - input_price: 0.14 - output_price: 0.58 - cache_read_price: 0.035 + input_price: 0.13 + output_price: 0.53 + cache_read_price: 0.033 max_output_tokens: 128000 supports_function_calling: true - name: zai-org/GLM-5.2