OpenAI's GPT-5.6 Family Arrives on Amazon Bedrock With Explicit Prompt Caching

OpenAI's GPT-5.6 model family — Sol, Terra, and Luna — is now generally available on Amazon Bedrock, paired with a new explicit prompt caching capability that gives developers manual control over which portions of a prompt get cached and reused across requests.

Three tiers, one API

GPT-5.6 ships in three capability tiers. Sol targets complex reasoning and agentic coding work. Terra is positioned for balanced, everyday production workloads. Luna handles fast, high-volume tasks like classification and summarization. All three are served through OpenAI's Responses API via Bedrock's bedrock-mantle endpoint, using model IDs openai.gpt-5.6-sol, openai.gpt-5.6-terra, and openai.gpt-5.6-luna.

Regional availability is limited at launch: Sol is available in US East (N. Virginia) and US East (Ohio). Terra and Luna add US West (Oregon). AWS says usage counts toward existing AWS spend commitments and runs under standard pay-per-token pricing, though exact per-token rates are not disclosed in the announcement — Bedrock's pricing page is the reference point, and AWS did not publish specific dollar figures for GPT-5.6 in this post.

GPT-5.6 supports five reasoning effort levels — none, low, medium, high, and xhigh — with medium as the default. According to OpenAI's Chris Dickens, who co-authored the announcement, the model is more token-efficient than GPT-5.5 and GPT-5.4, and AWS recommends teams migrating from earlier versions test one effort level lower than their current setting to capture savings without a quality loss. Sampling parameters like temperature and top_p only apply when reasoning effort is set to none; at other levels, the model's internal reasoning process governs output variance.

Explicit vs. implicit caching

Prompt caching on GPT-5.6 now runs in two modes. Implicit caching is on by default — Bedrock automatically places a cache breakpoint and reuses a stable prefix of at least 1,024 tokens with no code changes required. Explicit caching lets developers mark the cache boundary themselves, which AWS says produces higher hit rates for workloads where content after a stable system prompt or tool-definition prefix changes on every call, such as agentic tool loops and chat assistants.

The economics: cached reads are billed at a 90% discount versus uncached input tokens, while cache writes cost 1.25 times the standard uncached input rate. Cached content remains available for reuse for 30 minutes. Because a write costs more than a read saves, AWS states that a workload only turns a net cost benefit once cache reads account for roughly 20% of tokens flowing through the cache — a threshold explicit caching is designed to help workloads clear faster than the automatic implicit mode.

The feature targets agentic workflows specifically, where system instructions, tool schemas, and reference documents repeat across many sequential calls within a session.

What this means

This is primarily an infrastructure and cost-optimization update rather than a raw capability leap — AWS's announcement contains no new benchmark scores, context window figures, or parameter counts for GPT-5.6. The substantive news is twofold: OpenAI's newest frontier model family is now available inside AWS's governance and billing perimeter, and Bedrock customers running multi-step agentic pipelines get a direct lever to cut input-token costs. For teams already committed to AWS infrastructure, explicit caching removes one of the main cost objections to running frequent, prefix-heavy agent loops. The unanswered question is pricing transparency — without published per-token rates for GPT-5.6 on Bedrock, teams can't yet model total cost of ownership against calling OpenAI's API directly or competing hosted options.