OpenAI GPT-5.6 models launch on Amazon Bedrock with prompt caching

OpenAI's three new GPT-5.6 models—Sol, Terra, and Luna—are now generally available on Amazon Bedrock, OpenAI's managed endpoint on AWS. The models are served through the OpenAI-compatible Responses API on the bedrock-mantle endpoint. Sol is designed for the most complex reasoning and agentic coding work, Terra for balanced everyday production workloads, and Luna for fast, high-volume tasks such as classification and summarization. Customers access the models via short-term bearer tokens generated from AWS credentials, allowing code to remain free of long-lived secrets.

Alongside the model launch, OpenAI and AWS are introducing explicit prompt caching for GPT-5.6, a capability that gives developers precise control over which portions of a prompt are cached and reused across requests. Cached input tokens are billed at a 90 percent discount compared to uncached tokens, while cache writes are billed at 1.25 times the uncached input rate. Cached prefixes remain available for reuse for 30 minutes. The feature is designed for write-once, read-many patterns typical of agentic workflows, where system instructions, tool definitions, and reference documents repeat across many calls. A workload lowers its net input cost once cache reads make up roughly 20 percent of the tokens flowing through the cache.

Two caching modes are available. Implicit caching, the default, is enabled automatically with no code changes required; Amazon Bedrock places a cache breakpoint and tunes it for you. Explicit caching gives developers direct control by marking the exact end of a reusable prompt prefix through the prompt_cache_breakpoint parameter. A stable prefix must contain at least 1,024 tokens. Developers can set up to 4 breakpoints per request across input_text, input_image, and input_file blocks, and use prompt_cache_key to route requests to the same cache and prompt_cache_options to control caching mode and time-to-live. Every response reports cache behavior in the usage.input_tokens_details object, breaking down cached_tokens (read at a 90 percent discount) and cache_write_tokens (written at 1.25× the uncached rate).

GPT-5.6 supports adjustable reasoning effort levels—none, low, medium, high, and xhigh, with medium as the default. Higher levels allocate more reasoning to harder problems; none gives the lowest-latency path for straightforward tasks. The body recommends that teams upgrading from GPT-5.5 or GPT-5.4 start with the same effort level they use today, then test one level lower, since GPT-5.6 is more token-efficient and many workloads maintain quality at a lower effort setting. The models support streaming, function calling, and strict JSON schema output through the Responses API. Availability is limited to US regions: GPT-5.6 Sol is available in US East (N. Virginia) and US East (Ohio), while Terra and Luna are additionally available in US West (Oregon).