Oceans LLM

Appearance

OpenAI GPT-5.5 on Bedrock Mantle ​

See also: AWS Bedrock, Configuration Reference, Provider API Compatibility

This page shows how to configure OpenAI GPT-5.5 through Amazon Bedrock Mantle. It is for admins configuring gateway routes for users.

AWS lists GPT-5.5 with the following configuration:

  • model ID openai.gpt-5.5,
  • launch Region us-east-2,
  • endpoint kind bedrock-mantle,
  • responses path https://bedrock-mantle.{region}.api.aws/openai/v1/responses:

For more information and official documentation from AWS and OpenAI, see:

Bedrock API Key ​

Set AWS_BEARER_TOKEN_BEDROCK in the gateway runtime environment and configure bearer auth:

yaml
providers:
  - id: bedrock-mantle-openai
    type: aws_bedrock
    region: us-east-2
    endpoint_kind: bedrock_mantle
    auth:
      mode: bearer
      token: env.AWS_BEARER_TOKEN_BEDROCK
    display:
      label: Bedrock Mantle OpenAI
      icon_key: aws

models:
  - id: gpt-55-bedrock
    description: OpenAI GPT-5.5 on Amazon Bedrock Mantle
    tags: [bedrock, mantle, openai, gpt-5-5]
    routes:
      - provider: bedrock-mantle-openai
        upstream_model: openai.gpt-5.5
        capabilities:
          chat_completions: false
          responses: true
          stream: true
          embeddings: false
          tools: true
          vision: true
          json_schema: true
        compatibility:
          aws_bedrock:
            api_style: mantle_openai_responses
            openai_base_path: /openai/v1

For OpenAI-shaped Mantle routes, the gateway sends API-key auth as Authorization: Bearer ....

AWS SigV4 Provider Chain ​

Use auth.mode: default_chain when the gateway should sign Mantle requests with the AWS SDK default credential provider chain:

yaml
providers:
  - id: bedrock-mantle-openai
    type: aws_bedrock
    region: us-east-2
    endpoint_kind: bedrock_mantle
    auth:
      mode: default_chain

models:
  - id: gpt-55-bedrock
    routes:
      - provider: bedrock-mantle-openai
        upstream_model: openai.gpt-5.5
        capabilities:
          chat_completions: false
          responses: true
          stream: true
          embeddings: false
        compatibility:
          aws_bedrock:
            api_style: mantle_openai_responses
            openai_base_path: /openai/v1

For endpoint_kind: bedrock_mantle, SigV4 uses service name bedrock-mantle. For Runtime providers, SigV4 uses service name bedrock.

Bedrock Projects ​

Bedrock Projects for OpenAI-compatible APIs use the OpenAI-Project request header: AWS Projects. Configure it statically per route with extra_headers:

yaml
models:
  - id: gpt-55-bedrock
    routes:
      - provider: bedrock-mantle-openai
        upstream_model: openai.gpt-5.5
        capabilities:
          chat_completions: false
          responses: true
          stream: true
          embeddings: false
          tools: true
          vision: true
          json_schema: true
        extra_headers:
          OpenAI-Project: proj_123
        compatibility:
          aws_bedrock:
            api_style: mantle_openai_responses
            openai_base_path: /openai/v1

Projects apply to OpenAI-compatible Mantle routes. Do not rely on caller-supplied inbound headers for project routing; configure OpenAI-Project in the selected route.

Hosted Tool Compatibility ​

GPT-5.5 on Bedrock Mantle is configured as a Responses-first route through api_style: mantle_openai_responses. The tools: true capability in the examples means the route can attempt supported tool-bearing Responses workflows; it does not mean Bedrock supports every OpenAI-hosted tool.

Bedrock Mantle currently does not support the OpenAI-hosted image_generation Responses tool. Direct OpenAI GPT-5.5 can support hosted image generation, but the Bedrock OpenAI-compatible path has a narrower feature set.

Oceans strips opportunistic image_generation tool declarations before forwarding ordinary Bedrock-backed coding workflows. If a request explicitly requires image generation, for example through tool_choice: { "type": "image_generation" }, Oceans fails locally with a deterministic 400 instead of forwarding the request and exposing an opaque Bedrock validation error.

Revisit this shim when Bedrock Mantle supports hosted image_generation, or when Oceans grows explicit hosted-tool capabilities that can route image-generation-required requests to a provider that actually supports them.

For the broader rule, see Provider API Compatibility.

Caller Usage ​

Call the gateway's /v1/responses endpoint with the configured gateway model id:

json
{
  "model": "gpt-55-bedrock",
  "input": "Write a concise migration note."
}

Use /v1/chat/completions only for routes configured with a Chat Completions API style. GPT-5.5 on Bedrock is configured here as Responses-first with mantle_openai_responses.