OpenAI's GPT-5.6 lands on Bedrock, and the base URL is the whole story

The update

OpenAI GPT-5.6 Sol, Terra, and Luna are now generally available on Amazon Bedrock, accessed through the OpenAI Responses API on the new bedrock-mantle endpoint. Sol is the flagship reasoning model, Terra balances performance and cost for everyday production, and Luna is tuned for fast, low-cost, high-volume inference. All three share text and image input, a 272K-token context window, and selectable reasoning effort from none up to max. Pricing matches OpenAI's first-party rates and usage counts toward existing AWS commitments.

By the numbers

Under the hood

You hit the models at https://bedrock-mantle.{region}.api.aws with the Responses API served at /openai/v1/responses, using the OpenAI Python or TypeScript SDK. Authentication runs on an Amazon Bedrock API key or AWS credentials — the recommended path uses the BedrockOpenAI client with an auto-refreshing short-term token, since the environment-variable key expires within 12 hours. Every call runs under your IAM policies, inside your VPC, and lands in CloudTrail, with in-Region inference to help with data residency. OpenAI's naming now splits generation (the 5.6) from capability tier (Sol, Terra, Luna), so the tiers can advance on their own cadence, and switching between them needs no API changes.

The signal

The clever part isn't the models — it's that moving an existing OpenAI SDK app to Bedrock is mostly a base-URL swap and a model ID change. That lowers the switching cost to almost nothing, which is exactly how you win enterprises who don't want to run separate model infrastructure or babysit a second vendor relationship. The model race is quietly becoming a platform race: when frontier models show up inside the cloud you already trust, procurement, IAM, and CloudTrail logging start mattering as much as benchmarks.

The backstory

This fits a run of frontier labs meeting enterprises on the big cloud platforms rather than expecting migration to a separate stack. The bedrock-mantle endpoint and the openai/v1 path are OpenAI-specific plumbing bolted onto Bedrock — a sign AWS is willing to host the OpenAI SDK surface directly rather than forcing everyone onto Bedrock's native Converse API.

Who it's for

Developers

→Reuse OpenAI SDK code by swapping base URL

Enterprises

→OpenAI models under existing AWS controls and commitments

Founders

→No second vendor stack to run

The catch

Availability is thin: Sol runs only in US East (N. Virginia) and US East (Ohio), while Terra and Luna add US West (Oregon). These are third-party OpenAI models subject to OpenAI's terms, and classifier-flagged traffic is retained up to 30 days for automated abuse detection — retained inputs and outputs are stored and processed by AWS, not shared with the provider unless you opt in, and you control this through data retention mode. There are no published benchmarks here, so treat the tier descriptions as vendor positioning, not measured results.

What to watch

Watch for wider Regional rollout — Sol's two-Region limit is the obvious constraint for anyone outside US East. The Codex coding-agent integration and prompt-caching cost mechanics are where the practical payoff will show up, so the real test is how caching and quotas hold up under production agentic loads.