diff --git a/providers/aws-bedrock/global.openai.gpt-6.1-terra.yaml b/providers/aws-bedrock/global.openai.gpt-6.1-terra.yaml new file mode 100644 index 000000000..4df14f969 --- /dev/null +++ b/providers/aws-bedrock/global.openai.gpt-6.1-terra.yaml @@ -0,0 +1,480 @@ +canonical_model: openai/gpt-6-1-terra +costs: + - cache_creation_input_token_cost: 1.25e-06 + cache_read_input_token_cost: 5e-08 + input_cost_per_token: 1e-06 + output_cost_per_token: 5e-06 + region: us-west-2 + tiered_pricing: + cache_read: + - cost_per_token: 1e-07 + from: 272000 + cache_write: + - cost_per_token: 2.5e-06 + from: 272000 + input: + - cost_per_token: 2e-06 + from: 272000 + output: + - cost_per_token: 7.5e-06 + from: 272000 + pricing_mode: cumulative + - cache_creation_input_token_cost: 1.25e-06 + cache_read_input_token_cost: 5e-08 + input_cost_per_token: 1e-06 + output_cost_per_token: 5e-06 + region: us-west-1 + tiered_pricing: + cache_read: + - cost_per_token: 1e-07 + from: 272000 + cache_write: + - cost_per_token: 2.5e-06 + from: 272000 + input: + - cost_per_token: 2e-06 + from: 272000 + output: + - cost_per_token: 7.5e-06 + from: 272000 + pricing_mode: cumulative + - cache_creation_input_token_cost: 1.25e-06 + cache_read_input_token_cost: 5e-08 + input_cost_per_token: 1e-06 + output_cost_per_token: 5e-06 + region: us-east-2 + tiered_pricing: + cache_read: + - cost_per_token: 1e-07 + from: 272000 + cache_write: + - cost_per_token: 2.5e-06 + from: 272000 + input: + - cost_per_token: 2e-06 + from: 272000 + output: + - cost_per_token: 7.5e-06 + from: 272000 + pricing_mode: cumulative + - cache_creation_input_token_cost: 1.25e-06 + cache_read_input_token_cost: 5e-08 + input_cost_per_token: 1e-06 + output_cost_per_token: 5e-06 + region: us-east-1 + tiered_pricing: + cache_read: + - cost_per_token: 1e-07 + from: 272000 + cache_write: + - cost_per_token: 2.5e-06 + from: 272000 + input: + - cost_per_token: 2e-06 + from: 272000 + output: + - cost_per_token: 7.5e-06 + from: 272000 + pricing_mode: cumulative + - cache_creation_input_token_cost: 1.25e-06 + cache_read_input_token_cost: 5e-08 + input_cost_per_token: 1e-06 + output_cost_per_token: 5e-06 + region: eu-west-2 + tiered_pricing: + cache_read: + - cost_per_token: 1e-07 + from: 272000 + cache_write: + - cost_per_token: 2.5e-06 + from: 272000 + input: + - cost_per_token: 2e-06 + from: 272000 + output: + - cost_per_token: 7.5e-06 + from: 272000 + pricing_mode: cumulative + - cache_creation_input_token_cost: 1.25e-06 + cache_read_input_token_cost: 5e-08 + input_cost_per_token: 1e-06 + output_cost_per_token: 5e-06 + region: eu-west-1 + tiered_pricing: + cache_read: + - cost_per_token: 1e-07 + from: 272000 + cache_write: + - cost_per_token: 2.5e-06 + from: 272000 + input: + - cost_per_token: 2e-06 + from: 272000 + output: + - cost_per_token: 7.5e-06 + from: 272000 + pricing_mode: cumulative + - cache_creation_input_token_cost: 1.25e-06 + cache_read_input_token_cost: 5e-08 + input_cost_per_token: 1e-06 + output_cost_per_token: 5e-06 + region: eu-south-2 + tiered_pricing: + cache_read: + - cost_per_token: 1e-07 + from: 272000 + cache_write: + - cost_per_token: 2.5e-06 + from: 272000 + input: + - cost_per_token: 2e-06 + from: 272000 + output: + - cost_per_token: 7.5e-06 + from: 272000 + pricing_mode: cumulative + - cache_creation_input_token_cost: 1.25e-06 + cache_read_input_token_cost: 5e-08 + input_cost_per_token: 1e-06 + output_cost_per_token: 5e-06 + region: eu-south-1 + tiered_pricing: + cache_read: + - cost_per_token: 1e-07 + from: 272000 + cache_write: + - cost_per_token: 2.5e-06 + from: 272000 + input: + - cost_per_token: 2e-06 + from: 272000 + output: + - cost_per_token: 7.5e-06 + from: 272000 + pricing_mode: cumulative + - cache_creation_input_token_cost: 1.25e-06 + cache_read_input_token_cost: 5e-08 + input_cost_per_token: 1e-06 + output_cost_per_token: 5e-06 + region: eu-north-1 + tiered_pricing: + cache_read: + - cost_per_token: 1e-07 + from: 272000 + cache_write: + - cost_per_token: 2.5e-06 + from: 272000 + input: + - cost_per_token: 2e-06 + from: 272000 + output: + - cost_per_token: 7.5e-06 + from: 272000 + pricing_mode: cumulative + - cache_creation_input_token_cost: 1.25e-06 + cache_read_input_token_cost: 5e-08 + input_cost_per_token: 1e-06 + output_cost_per_token: 5e-06 + region: eu-central-2 + tiered_pricing: + cache_read: + - cost_per_token: 1e-07 + from: 272000 + cache_write: + - cost_per_token: 2.5e-06 + from: 272000 + input: + - cost_per_token: 2e-06 + from: 272000 + output: + - cost_per_token: 7.5e-06 + from: 272000 + pricing_mode: cumulative + - cache_creation_input_token_cost: 1.25e-06 + cache_read_input_token_cost: 5e-08 + input_cost_per_token: 1e-06 + output_cost_per_token: 5e-06 + region: eu-central-1 + tiered_pricing: + cache_read: + - cost_per_token: 1e-07 + from: 272000 + cache_write: + - cost_per_token: 2.5e-06 + from: 272000 + input: + - cost_per_token: 2e-06 + from: 272000 + output: + - cost_per_token: 7.5e-06 + from: 272000 + pricing_mode: cumulative + - cache_creation_input_token_cost: 1.25e-06 + cache_read_input_token_cost: 5e-08 + input_cost_per_token: 1e-06 + output_cost_per_token: 5e-06 + region: ca-central-1 + tiered_pricing: + cache_read: + - cost_per_token: 1e-07 + from: 272000 + cache_write: + - cost_per_token: 2.5e-06 + from: 272000 + input: + - cost_per_token: 2e-06 + from: 272000 + output: + - cost_per_token: 7.5e-06 + from: 272000 + pricing_mode: cumulative + - cache_creation_input_token_cost: 1.25e-06 + cache_read_input_token_cost: 5e-08 + input_cost_per_token: 1e-06 + output_cost_per_token: 5e-06 + region: ap-southeast-7 + tiered_pricing: + cache_read: + - cost_per_token: 1e-07 + from: 272000 + cache_write: + - cost_per_token: 2.5e-06 + from: 272000 + input: + - cost_per_token: 2e-06 + from: 272000 + output: + - cost_per_token: 7.5e-06 + from: 272000 + pricing_mode: cumulative + - cache_creation_input_token_cost: 1.25e-06 + cache_read_input_token_cost: 5e-08 + input_cost_per_token: 1e-06 + output_cost_per_token: 5e-06 + region: ap-southeast-6 + tiered_pricing: + cache_read: + - cost_per_token: 1e-07 + from: 272000 + cache_write: + - cost_per_token: 2.5e-06 + from: 272000 + input: + - cost_per_token: 2e-06 + from: 272000 + output: + - cost_per_token: 7.5e-06 + from: 272000 + pricing_mode: cumulative + - cache_creation_input_token_cost: 1.25e-06 + cache_read_input_token_cost: 5e-08 + input_cost_per_token: 1e-06 + output_cost_per_token: 5e-06 + region: ap-southeast-5 + tiered_pricing: + cache_read: + - cost_per_token: 1e-07 + from: 272000 + cache_write: + - cost_per_token: 2.5e-06 + from: 272000 + input: + - cost_per_token: 2e-06 + from: 272000 + output: + - cost_per_token: 7.5e-06 + from: 272000 + pricing_mode: cumulative + - cache_creation_input_token_cost: 1.25e-06 + cache_read_input_token_cost: 5e-08 + input_cost_per_token: 1e-06 + output_cost_per_token: 5e-06 + region: ap-southeast-4 + tiered_pricing: + cache_read: + - cost_per_token: 1e-07 + from: 272000 + cache_write: + - cost_per_token: 2.5e-06 + from: 272000 + input: + - cost_per_token: 2e-06 + from: 272000 + output: + - cost_per_token: 7.5e-06 + from: 272000 + pricing_mode: cumulative + - cache_creation_input_token_cost: 1.25e-06 + cache_read_input_token_cost: 5e-08 + input_cost_per_token: 1e-06 + output_cost_per_token: 5e-06 + region: ap-southeast-2 + tiered_pricing: + cache_read: + - cost_per_token: 1e-07 + from: 272000 + cache_write: + - cost_per_token: 2.5e-06 + from: 272000 + input: + - cost_per_token: 2e-06 + from: 272000 + output: + - cost_per_token: 7.5e-06 + from: 272000 + pricing_mode: cumulative + - cache_creation_input_token_cost: 1.25e-06 + cache_read_input_token_cost: 5e-08 + input_cost_per_token: 1e-06 + output_cost_per_token: 5e-06 + region: ap-southeast-1 + tiered_pricing: + cache_read: + - cost_per_token: 1e-07 + from: 272000 + cache_write: + - cost_per_token: 2.5e-06 + from: 272000 + input: + - cost_per_token: 2e-06 + from: 272000 + output: + - cost_per_token: 7.5e-06 + from: 272000 + pricing_mode: cumulative + - cache_creation_input_token_cost: 1.25e-06 + cache_read_input_token_cost: 5e-08 + input_cost_per_token: 1e-06 + output_cost_per_token: 5e-06 + region: ap-south-1 + tiered_pricing: + cache_read: + - cost_per_token: 1e-07 + from: 272000 + cache_write: + - cost_per_token: 2.5e-06 + from: 272000 + input: + - cost_per_token: 2e-06 + from: 272000 + output: + - cost_per_token: 7.5e-06 + from: 272000 + pricing_mode: cumulative + - cache_creation_input_token_cost: 1.25e-06 + cache_read_input_token_cost: 5e-08 + input_cost_per_token: 1e-06 + output_cost_per_token: 5e-06 + region: ap-northeast-3 + tiered_pricing: + cache_read: + - cost_per_token: 1e-07 + from: 272000 + cache_write: + - cost_per_token: 2.5e-06 + from: 272000 + input: + - cost_per_token: 2e-06 + from: 272000 + output: + - cost_per_token: 7.5e-06 + from: 272000 + pricing_mode: cumulative + - cache_creation_input_token_cost: 1.25e-06 + cache_read_input_token_cost: 5e-08 + input_cost_per_token: 1e-06 + output_cost_per_token: 5e-06 + region: ap-northeast-1 + tiered_pricing: + cache_read: + - cost_per_token: 1e-07 + from: 272000 + cache_write: + - cost_per_token: 2.5e-06 + from: 272000 + input: + - cost_per_token: 2e-06 + from: 272000 + output: + - cost_per_token: 7.5e-06 + from: 272000 + pricing_mode: cumulative + - cache_creation_input_token_cost: 1.25e-06 + cache_read_input_token_cost: 5e-08 + input_cost_per_token: 1e-06 + output_cost_per_token: 5e-06 + region: ap-east-2 + tiered_pricing: + cache_read: + - cost_per_token: 1e-07 + from: 272000 + cache_write: + - cost_per_token: 2.5e-06 + from: 272000 + input: + - cost_per_token: 2e-06 + from: 272000 + output: + - cost_per_token: 7.5e-06 + from: 272000 + pricing_mode: cumulative + - cache_creation_input_token_cost: 1.25e-06 + cache_read_input_token_cost: 5e-08 + input_cost_per_token: 1e-06 + output_cost_per_token: 5e-06 + region: af-south-1 + tiered_pricing: + cache_read: + - cost_per_token: 1e-07 + from: 272000 + cache_write: + - cost_per_token: 2.5e-06 + from: 272000 + input: + - cost_per_token: 2e-06 + from: 272000 + output: + - cost_per_token: 7.5e-06 + from: 272000 + pricing_mode: cumulative +features: + - function_calling + - parallel_function_calling + - tool_choice + - system_messages + - json_output + - prompt_caching + - structured_output +limits: + context_window: 1050000 + max_input_tokens: 922000 + max_output_tokens: 128000 + max_tokens: 128000 +modalities: + input: + - text + - image + output: + - text +mode: chat +model: global.openai.gpt-6.1-terra +params: + - defaultValue: medium + key: reasoning_effort + supportedValues: + - none + - low + - medium + - high + - xhigh + - max + type: string +provisioning: serverless +sources: + - https://docs.aws.amazon.com/bedrock/latest/userguide/model-cards-openai.html + - https://developers.openai.com/api/docs/models/gpt-6.1-terra +status: active +supportedModes: + - chat + - responses +thinking: true