Introducing explicit prompt caching for OpenAI GPT-5.6 models on Amazon Bedrock
AWS Melanie Li ● Covered by 2 sources
OpenAI's GPT-5.6 model family (Sol, Terra, Luna) just landed on Amazon Bedrock. The real news is caching: reused prompt chunks now cost 90% less for AI agents.
Based on reporting by AWS, Melanie Li — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Amazon Bedrock just picked up OpenAI's newest model lineup, and it's arriving as a full family rather than a single release. GPT-5.6 Sol is built for the hardest reasoning and agentic coding jobs, Terra is the workhorse for everyday production traffic, and Luna handles fast, high-volume work like classification and summarization. All three run through the same OpenAI-compatible Responses API on Bedrock's bedrock-mantle endpoint, billed per token, wrapped in AWS's usual security and governance layer, and counted against whatever AWS commitments a customer already has. Sol is live in US East (N. Virginia) and US East (Ohio); Terra and Luna also run in US West (Oregon).
But the feature getting the real attention here isn't the models themselves — it's explicit prompt caching. Anyone who's built an agent knows the pain: the same system instructions, tool definitions, and reference documents get shipped over and over with every single call, and you pay full price each time. GPT-5.6 lets developers mark exactly where a reusable prompt prefix ends, so everything before that point gets cached and everything after it can change freely. Cached reads come at a 90 percent discount versus uncached input, and the cache sticks around for reuse for at least 30 minutes — plenty of time to cover a burst of calls inside one agent run.
There's a tradeoff baked into the pricing that's easy to miss: writing to the cache actually costs 1.25 times the normal input rate. So caching only pays off once a workload's cache-read share climbs to roughly 20 percent of the tokens flowing through it. Below that, you're paying a premium for nothing. Implicit caching, which Bedrock turns on automatically for any stable prefix of at least 1,024 tokens, gets you partway there without touching your code. Explicit mode — where developers plant up to four breakpoints per request across text, image, or file blocks — pushes the hit rate higher because it removes the guesswork about where the reusable content actually ends.
The numbers AWS shows for a support-assistant example make the case plainly. A 3,626-token system prompt gets written to cache on the first call. From the second call onward, that entire block reads back at the discounted rate while only the fresh user question — a few dozen tokens — gets billed at standard rates. Total input token counts barely move call to call, even though the underlying prompt is thousands of tokens long each time.
Worth noting too: GPT-5.6 lets developers dial reasoning effort from none up through xhigh, with medium as the default, and AWS is nudging anyone migrating from GPT-5.5 or GPT-5.4 to try one effort level lower than they're used to, since the new generation is described as more token-efficient. Between tiered models, tunable reasoning, and caching that specifically targets the repetitive backbone of agentic workflows, this reads less like a model drop and more like Amazon and OpenAI building out the plumbing that makes running agents at scale actually affordable.
My take — AI-written commentary, not fact-checked reporting
Nobody launches three new model tiers and expects the pricing mechanic buried in paragraph six to be the headline, but it should be here. A 90 percent discount on repeated prompt content is the kind of unglamorous infrastructure work that actually determines whether agentic AI is viable at scale, far more than another benchmark chart ever will. The catch that cache writes cost more than standard input is a useful reality check too — caching isn't free magic, it's a bet that only pays off once your agent is chatty enough to reuse that prefix again and again. That's a very specific kind of company: one already running agents in production, not one still experimenting.}
Read more about this at: AWS