The Vibe Father
Amazon Bedrock now supports explicit prompt caching for OpenAI GPT‑5.6 Sol, Terra, and Luna. Developers can mark the stable part of a prompt for reuse instead of relying only on automatically chosen cache boundaries. AWS says cached input reads receive a 90 percent discount and remain reusable for 30 minutes.
Where caching helps
Agent loops often repeat system instructions, tool definitions, repository guidance, and reference documents. Caching lets Bedrock avoid recomputing that stable prefix on later calls. The best workload writes a large prefix once and reads it many times while the task-specific tail changes.
| Prompt content | Placement |
|---|---|
| Stable system rules | Before the cache boundary |
| Deterministically ordered tool definitions | Before the boundary |
| Shared reference material | Before the boundary if reused |
| User request and current state | After the boundary |
| Timestamps and request IDs | After the boundary |
The cost math
AWS says cache writes cost 1.25 times the normal input rate and reads cost ten percent of the normal input rate. Because writing costs more, caching a prefix that is never reused loses money. AWS estimates the workload reduces net input cost once cache reads represent roughly 20 percent of tokens flowing through the cache.
The 30-minute reuse window also matters. A prefix used daily will be rewritten each session. A prefix reused across a burst of agent turns can produce a much better hit rate.
Bedrock API details
The GPT‑5.6 family is served through an OpenAI-compatible Responses API on the Bedrock Mantle endpoint. AWS recommends short-term bearer tokens derived from AWS credentials instead of long-lived API keys. Sol is listed in two US East regions, while Terra and Luna are also available in US West Oregon.
How to verify the gain
- Track cache write, read, and uncached input tokens separately.
- Keep tool ordering stable so identical prefixes remain identical.
- Move volatile values behind the cache boundary.
- Measure accepted-task cost, not only the input discount.
- Confirm that stale cached instructions cannot outlive a critical policy change.
Bottom line
Explicit prompt caching is a real cost and latency lever for repeated agent context. It is not free by default. Design a stable prefix, reuse it within the window, measure the hit rate, and invalidate it deliberately when instructions or tools change.
