AWS brings Kimi Highspeed to Bedrock with separate cache pricing AWS brings Kimi Highspeed to Bedrock with separate cache pricing

What did AWS announce?

The addition gives Bedrock customers another route to Moonshot’s model for coding and document work. AWS describes it as an open-weight, multimodal model with a million-token context window. This is a hosting announcement, not a newly trained model. AWS’s documentation history records the model-card addition on November 4, two days before the public announcement. The distinction matters when dating the release: an earlier documentation entry and a later announcement describe the same hosted offering.

What does the input bill measure?

A cached input reuses processing from an earlier request. Ordinary input, writing reusable context and reading it again occupy different billing categories. AWS’s caching guide explains that a stable prefix can serve later questions about the same material. A changed prefix produces a miss. This makes prompt structure part of the economics: repeated instructions or reference material have a different reuse pattern from a succession of unrelated questions. The amount submitted alone does not identify which category the service will bill.

How does the arithmetic work?

Consider a hypothetical workload that sends the same 100,000-token reference prefix ten times. Ignore changing question text and all output, so the example isolates that prefix. At the Global Standard rates, ten ordinary input charges would total $4. One cache write costs $0.50. If the remaining nine requests each read the entire prefix from cache, those reads total $0.18. The combined prefix cost is $0.68, a saving of $3.32 against the ordinary-input baseline. These are calculated scenarios, not measured performance or a promise of the bill a customer will receive.

What happens when reuse fails?

Now keep the same ten requests but assume five cache writes and five reads. The writes cost $2.50 and the reads cost $0.10, producing $2.60 for the prefix. If all ten requests instead write new entries, the prefix costs $5. That exceeds the $4 ordinary-input baseline by $1. The price of one read is therefore a poor stand-in for the price of a session. The relevant quantity is the mixture of writes, reads and ordinary input across completed work. Output remains a separate charge in every version of this example.

What changes outside the calculation?

AWS’s guide describes a cache entry as belonging to the server that processed the request. Routing a later request elsewhere can produce another write. Concurrent requests can also arrive before an identical request has begun responding. Explicit cache controls select reusable material; they do not reserve a future hit. The launch also enters an existing hosting market: Runpod’s November 6 guide offers both Pro and Highspeed through its public endpoint. That choice makes the hosting interface and billing behavior part of integration work, alongside the model’s answers.