feat(bedrock): add OpenAI GPT-5.6 (Sol/Terra/Luna) support via bedrock-mantle endpoint

[jeremyluojl]

Summary

• Add GPT-5.6 Sol, GPT-5.6 Terra, and GPT-5.6 Luna as selectable models in the Amazon Bedrock plugin
• Route these models through the bedrock-mantle endpoint using the OpenAI Responses API, reusing the infrastructure introduced in feat(bedrock): add OpenAI GPT-5.5 and GPT-5.4 support via bedrock-mantle endpoint #3240 for GPT-5.5 / GPT-5.4
• Add a Reasoning Effort parameter (none / low / medium / high / xhigh / max) forwarded to the Responses API reasoning.effort
• Do not forward top_p on the bedrock-mantle path — GPT-5.5 / GPT-5.6 reject it, and it is incompatible with reasoning even for GPT-5.4

Background

GPT-5.6 (Sol / Terra / Luna) became generally available on Amazon Bedrock on July 13, 2026. Like GPT-5.5 / GPT-5.4 (added in #3240), these models use the bedrock-mantle endpoint with the OpenAI Responses API rather than the bedrock-runtime Converse API used by GPT OSS, Claude, etc.

Model	Positioning	Bedrock model ID
GPT-5.6 Sol	Most capable (frontier reasoning / agentic)	openai.gpt-5.6-sol
GPT-5.6 Terra	Balanced, competitive with GPT-5.5 at ~half the cost	openai.gpt-5.6-terra
GPT-5.6 Luna	Lightweight / lowest cost	openai.gpt-5.6-luna

All three are Responses-API-only, bedrock-mantle-only, In-Region only (no cross-region inference), with a 272K-token context window — consistent with the GPT-5.5 / GPT-5.4 path.

Changes

models/bedrock/models/llm/model_ids.py

• Add GPT-5.6 Sol / GPT-5.6 Terra / GPT-5.6 Luna → openai.gpt-5.6-sol / -terra / -luna to the openai section
• Add the three model IDs to the is_support_cross_region exclusion list (In-Region only)

models/bedrock/models/llm/openai.yaml

• Add the three GPT-5.6 models to the model_name options
• Add a reasoning_effort parameter rule (none / low / medium / high / xhigh / max, default medium)

models/bedrock/models/llm/llm.py

• Add the three model IDs to _BEDROCK_MANTLE_MODEL_IDS
• Add them to the pricing name mapping in _get_model_specific_pricing
• Forward reasoning_effort to the Responses API as reasoning={"effort": ...}
• Stop forwarding top_p on the bedrock-mantle path (see Known Limitations)
• Extend the resolved_model_name fallback to cover the GPT-5.6 variants

models/bedrock/models/llm/model_configurations/gpt-5-6-{sol,terra,luna}.yaml

• New per-model pricing/config files (context 272K). Pricing is the OpenAI list price for In-Region on-demand inference on Bedrock (unit: 0.001 per 1K tokens):
  • Sol: input 0.0055 / output 0.033
  • Terra: input 0.00275 / output 0.0165
  • Luna: input 0.0011 / output 0.0066

models/bedrock/manifest.yaml

• Bump plugin version 0.0.74 → 0.0.75

No dependency changes: openai and aws-bedrock-token-generator were already added in #3240.

Known Limitations

• top_p not forwarded: GPT-5.5 / GPT-5.6 reject top_p (400 unsupported_parameter), and top_p is mutually exclusive with reasoning even on GPT-5.4. Since these are reasoning models driven by reasoning_effort, top_p is omitted on the mantle path. It remains in openai.yaml only for the GPT OSS Converse models (separate code path, unaffected).
• Cross-region inference: Not available (In-Region only), per AWS documentation.
• Regional availability (verified via the bedrock-mantle /v1/models endpoint):
  • us-east-1, us-east-2: all three (Sol / Terra / Luna)
  • us-west-2: Terra and Luna only (Sol not available)
• Tool use / image input / structured output: not supported on this path (same as feat(bedrock): add OpenAI GPT-5.5 and GPT-5.4 support via bedrock-mantle endpoint #3240).

References

• GPT-5.6 Sol model card
• GPT-5.6 Terra model card
• GPT-5.6 Luna model card
• Builds on feat(bedrock): add OpenAI GPT-5.5 and GPT-5.4 support via bedrock-mantle endpoint #3240 (GPT-5.5 / GPT-5.4 bedrock-mantle support)