aws: OpenAI GPT-6 Sol/Luna/Astra and GPT-5.6 on Bedrock fail every turn with a 400 when temperature or top_p is set

[kimnamu]

Bug Description

Thanks for the quick review and merge of #7411. The _MODELS_REJECTING_SAMPLING_PARAMS guard from #6947 made that fix small, and I think the same list is all that's needed here.

OpenAI GPT-6 Sol and GPT-6 Luna became available on Bedrock on 2026-09-22, and GPT-6 Astra on 2026-09-08. Converse rejects inferenceConfig.temperature and topP for all three, and for GPT-5.6, at any value (including 1). The guard only lists Claude models, so aws.LLM(temperature=...) or top_p=... sends both fields and every turn ends in a 400:

aws.LLM(model=..., temperature=0.7, top_p=0.9), us-east-1	main (b4df92e)
us.openai.gpt-6-sol / -luna / -astra	❌ APIConnectionError after 4 attempts (about 7 s)
global.openai.gpt-6-sol / -luna / -astra	❌ same
us.openai.gpt-6-astra with only temperature=1.0, or only top_p=1.0	❌ same
us.openai.gpt-5.6-sol	❌ same
openai.gpt-oss-120b-1:0	✅ pong (accepts both fields)
GPT-6 without temperature/top_p	✅ pong

Expected Behavior

Same as Claude Opus 4.7+: both fields are dropped, with the existing one-time warning, and the turn succeeds. gpt-oss keeps receiving them.

Reproduction Steps

cat > repro.py <<'PY'
import asyncio
from livekit.agents.llm import ChatContext
from livekit.plugins import aws

async def main():
    for model in ("us.openai.gpt-6-sol", "us.openai.gpt-6-luna", "us.openai.gpt-6-astra",
                  "openai.gpt-oss-120b-1:0"):
        ctx = ChatContext()
        ctx.add_message(role="user", content="Say pong and nothing else.")
        try:
            r = await aws.LLM(model=model, temperature=0.7, top_p=0.9).chat(chat_ctx=ctx).collect()
            print(model, "->", repr(r.text))
        except Exception as e:
            print(model, "->", type(e).__name__, e)

asyncio.run(main())
PY
uv run python repro.py

us.openai.gpt-6-sol -> APIConnectionError failed to generate LLM completion after 4 attempts (caused by ValidationException: An error occurred (ValidationException) when calling the ConverseStream operation: This model doesn't support the temperature field. Remove temperature and try again.)
us.openai.gpt-6-luna -> APIConnectionError (same)
us.openai.gpt-6-astra -> APIConnectionError (same)
openai.gpt-oss-120b-1:0 -> 'pong'

With only top_p=0.9: This model doesn't support the topP field. Remove topP and try again.

Operating System

macOS 26 (the rejection is server-side, so any OS)

Models Used

AWS Bedrock us./global. openai.gpt-6-sol, openai.gpt-6-luna, openai.gpt-6-astra, openai.gpt-5.6-sol via livekit-plugins-aws aws.LLM

Package Versions

livekit-agents==1.8.2
livekit-plugins-aws==1.8.2   # main at b4df92e
aiobotocore==3.8.0
botocore==1.43.46
python==3.13

Session/Room/Call IDs

N/A, reproduces with aws.LLM alone.

Proposed Solution

_MODELS_REJECTING_SAMPLING_PARAMS = (
    ...,
    "claude-fable-5",
    # OpenAI GPT-5.6/GPT-6: "This model doesn't support the temperature field. Remove
    # temperature and try again." gpt-oss still accepts both, so it stays off this list.
    "gpt-5.6",
    "gpt-6",
)

The substrings don't match gpt-oss-*. Application inference profiles stay on supports_sampling_params, as today. I have this with hermetic tests and can open a PR.

Additional Context

Scope and workaround

• Workaround today: aws.LLM(..., supports_sampling_params=False). The request goes out with inferenceConfig: {} and the turn succeeds.
• Other GPT-6 paths through aws.LLM worked for all three models: plain, system prompt, streaming, tool round trip, tool_choice required/named, image input, usage, global. profiles.
• The 400 is retried three times before it surfaces because nothing was streamed yet. That is a separate problem and not part of this issue.