Bedrock's cache write on the first request pulls the 90 percent discount down to 75

Take the ten-question example and price it in units of one full-price pass over the document. Ten uncached requests cost 10 units. With caching, the first request writes the cache and the next nine read it at 90 percent off, so the nine reads cost 0.9 units ^[1][3]. If the write were billed at the plain input rate, the total would be 1.9 against 10, a saving of 81 percent ^[2]. AWS puts the net saving at about 75 percent ^[3]. Back the write out of that figure: 0.25 x 10 = 2.5 units, less 0.9 for the reads, leaves roughly 1.6 units for one write, or about 1.6 times a standard input token ^[1]. The post links to the Bedrock pricing page for the actual rates ^[10]. If the real write multiplier is nearer 1.25, the saving beats AWS's own number.

Add questions and the write amortises. Without caching, AWS says a 10,000-token contract sent alongside 50 user questions bills 500,000 input tokens at full price ^[2]. Cached, that is one write plus 49 reads: 1.6 + 4.9 = 6.5 units against 50, an 87 percent saving ^[3]. The 90 percent applies to cached input tokens on a hit, and cache reads and writes are categories on top of standard input and output tokens, which still bill as usual ^[1][10].

Two settings decide whether any of it activates, and neither is in your prompt text. Each checkpoint must clear a token minimum: 1,024 tokens for Anthropic Claude Sonnet 4.5 and Sonnet 4.6, and 4,096 for Opus models, four times as much ^[5][4]. A 600-token system prompt with a cachePoint on it caches nothing on Sonnet 4.5 ^[5].

The second is ordering. Bedrock evaluates whether the content preceding the cachePoint marker matches an existing entry; a hit skips reprocessing and generation starts from the cached state, a miss processes the full content and writes it ^[4]. Anything that varies per request therefore has to sit after the marker. Put a session id or a timestamp above it and every request is a miss ^[6].

TTL is five minutes by default, with select models supporting up to an hour, and a request after expiration triggers a new write ^[6][7]. Five minutes is not long for a person reading a contract between questions. Entries are also scoped to individual AWS accounts and Regions ^[8], so an active-active deployment across two Regions warms two entries and pays two writes ^[7].

On the alternatives, the post is straight about the cost of each: shortened prompts "might also reduce context quality", smaller context windows lower cost "at the expense of the model's ability to reason over complete information", and response caching "provides no benefit when the same context is paired with different questions" ^[11]. Caching moves the decision into the request body, which the post says reduces time-to-first-token and cached input cost "without changing your model or prompt quality" ^[12]. The cachePoint syntax is identical across supported families including Anthropic Claude and Amazon Nova ^[9]. The six walkthroughs use Claude Sonnet 4.5 and need model access in a supported Region such as us-west-2 ^[13].