feat(openai): add LLM.with_aws_bedrock for OpenAI models on Amazon Bedrock

[piyush-gambhir]

Summary

Adds support for OpenAI models served through Amazon Bedrock, mirroring the existing LLM.with_azure(...) integration. Bedrock's OpenAI-compatible bedrock-mantle endpoint serves two families with different API surfaces and URL paths, so this adds a with_aws_bedrock(...) factory to both the Chat Completions and Responses LLMs:

Models	API on Bedrock	Mantle path	Factory
gpt-oss-120b, gpt-oss-20b	Chat Completions (+ Responses)	/v1	openai.LLM.with_aws_bedrock(...)
gpt-5.5, gpt-5.4	Responses only	/openai/v1	openai.responses.LLM.with_aws_bedrock(...)

Both build an openai.AsyncBedrockOpenAI client and resolve the regional endpoint automatically. The openai SDK only ever derives the /openai/v1 path (which serves the gpt-5.x models); the gpt-oss models live on /v1, so the correct path is resolved from the model id. For the Responses variant the WebSocket transport is disabled automatically, since Bedrock only exposes the HTTP Responses path.

Usage

from livekit.plugins import openai

# Chat Completions — gpt-oss
llm = openai.LLM.with_aws_bedrock(model="openai.gpt-oss-120b", aws_region="us-east-2")

# Responses API — gpt-5.5 / gpt-5.4
llm = openai.responses.LLM.with_aws_bedrock(model="openai.gpt-5.5", aws_region="us-east-2")

# Refreshable credentials (e.g. aws-bedrock-token-generator)
from aws_bedrock_token_generator import provide_token
llm = openai.responses.LLM.with_aws_bedrock(
    model="openai.gpt-5.5", bedrock_token_provider=provide_token, aws_region="us-east-2",
)

api_key, aws_region, and base_url are inferred from AWS_BEARER_TOKEN_BEDROCK, AWS_REGION/AWS_DEFAULT_REGION, and AWS_BEDROCK_BASE_URL respectively when not passed explicitly. api_key and bedrock_token_provider are mutually exclusive.

Model IDs: the mantle (OpenAI-compatible) endpoint uses IDs without the -1:0 suffix (e.g. openai.gpt-oss-120b). The -1:0 form is only for the bedrock-runtime InvokeModel/Converse APIs.

Changes

• LLM.with_aws_bedrock(...) on the Chat Completions LLM (gpt-oss)
• responses.LLM.with_aws_bedrock(...) on the Responses LLM (gpt-5.5, gpt-5.4, gpt-oss)
• model-aware mantle path resolution (/v1 for gpt-oss, /openai/v1 for gpt-5.x)
• BedrockChatModels / BedrockResponsesModels literals
• AsyncBedrockTokenProvider type alias for refreshable credentials
• Bump openai to >=2.40 — the first release shipping AsyncBedrockOpenAI
• README + module docstring mention Amazon Bedrock
• Hermetic unit tests (tests/test_openai_bedrock.py)

Testing

• ruff format / ruff check — clean
• mypy (strict) on livekit.plugins.openai — clean
• pytest tests/test_openai_bedrock.py --unit — 5 passed
• Verified end-to-end against live Bedrock (us-east-2): gpt-5.5 and gpt-5.4 (Responses) and gpt-oss-120b (Chat Completions) all stream completions through the wrappers.

The Chat Completions variant uses _strict_tool_schema=False, consistent with the other open-weight (gpt-oss) OpenAI-compatible providers in this plugin (Cerebras, SambaNova). Model availability/paths/regions confirmed against the AWS model cards for gpt-oss-120b, GPT-5.5, and GPT-5.4.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 4 additional findings.