Bedrock's two GPT-6 Astra endpoints split prompt caching from invocation logging

The split follows the control plane. Guardrails exist only on bedrock-runtime, and only through Converse ^[7]. Runtime also serves the Responses and Chat Completions APIs ^[5]. Those calls are synchronous only, and a request with background set returns a 400 ^[8]. Converse and Responses calls on runtime are captured by invocation logging to S3 or CloudWatch, with 100 KB inline ^[9]. bedrock-mantle exposes the OpenAI-shaped paths, /openai/v1/responses and /openai/v1/chat/completions ^[6], with async calls and server-side tools ^[10].

The cache is the well-built part. Breakpoints are explicit: a request sets up to four with `prompt_cache_breakpoint`, each cached prefix needs at least 1,024 tokens, and entries expire after 30 minutes ^[11]. A reviewer can see in the code which prefix is reused and for how long. Reads cost $1.10 per million tokens ^[12], one-tenth of the $11 in-Region input rate ^[23].

That discount is paid for in audit coverage. Mantle requests never reach the invocation log. CloudTrail records the bedrock-mantle:CreateInference action and its caller, without the content ^[13]. Separately, Responses calls default to `store=true`, keeping input and output for 30 days, and a call routed through the global profile is stored in the Region that served it ^[14]. The global profile can route to any commercial Region ^[16].

Existing clients break first. A wrapper standardized on InvokeModel to get one instrumentation point does not call Astra ^[5]. CountTokens is marked unsupported on bedrock-runtime for this model, as are structured outputs ^[20]. Above 272,000 input tokens the whole request reprices, not just the overflow ^[18]. A 280,000-token prompt on global costs $5.60 in input, and the same prompt trimmed to 270,000 costs $2.70 ^[19]. The prompt grew 3.7% and the input bill grew 107% ^[24]. Bedrock drew a price line at 272,000 tokens and left out the API that counts them, so any pre-flight check has to run on the client ^[20].

Capacity has one setting. Standard is the only supported tier. Priority, Flex and Reserved are not, so Region is the only control over tail latency ^[15]. The us. profile covers five Regions in the US and Canada ^[16]. Geographic and in-Region routing cost $11 per million input tokens against $10 on global ^[17], a 10% premium for knowing where inference runs ^[26].

In my context, a regulated workload where Guardrails are mandatory, I would put the audited path on runtime Converse with a geographic profile and pay the uncached $11 per million input tokens ^[17]. A cache-heavy internal agent goes to mantle, with request and response capture built into the application, because no Bedrock log will hold that content ^[13]. The lifecycle sets how long that split has to hold: end-of-life no sooner than September 8, 2027, with a legacy period of at least six months ^[21]. GA to the earliest end-of-life is twelve months ^[25]. The author of the dev.to analysis wrote that this is "a one-year contractual floor for planning migration, and it is the kind of fact that belongs in your ADR, not in your slide" ^[22].