The Heavy-Workload Team’s Claude Sonnet 5.5 Cache Bill

01 · Step 1: Separate the prompt from the answer

Start with the shape of a typical request. A large engineering workflow may send a stable system instruction, repository guidance, tool definitions, project documentation, and a short task-specific question. The stable portion can often be written to a provider cache. The changing portion usually cannot.

The important distinction is that the model processes both portions, but the billing categories may differ. Input tokens that are newly processed are ordinary input tokens. Tokens stored for later reuse are cache write tokens. Tokens reused from an existing cache entry are cache read tokens. Output tokens remain a separate category.

02 · Step 2: Identify the repeatable prefix

Review the request sequence and mark what stays unchanged between calls. A team using Claude Sonnet 5.5 for repository work might send the same codebase instructions and tool schema across hundreds of requests, while changing only the file list, user request, and recent tool results.

Cache efficiency depends on this stable prefix. If a changing message appears before the reusable material, the cache may be less useful or may fail to match the intended prefix. Keep durable instructions and shared context together, then place task-specific content after them.

03 · Step 3: Treat the first request as a write event

The first request that establishes a cacheable prefix generally creates a cache write event. That write is not the same as a cache read. The system has to process and store the prefix before later calls can reuse it.

At scale, writes can be easy to overlook because they happen less often than reads. They still belong in the bill. A deployment that refreshes its cache frequently, changes the prefix on every request, or uses short cache lifetimes can produce many write events and reduce the expected savings from reuse.

04 · Step 4: Count every reuse as a read

Once the prefix matches an active cache entry, later calls can generate cache read tokens. A read is charged according to the relay and model pricing rules for that model; it is not automatically free simply because the text was sent before.

This is the central budgeting point. A large team may send the same context repeatedly and see most of its input volume reported as cache reads. That can be more efficient than repeatedly sending the same material as ordinary input, but the read volume still grows with request count and cached prefix length.

05 · Step 5: Calculate the bill from metered categories

Build a simple usage table with four rows: ordinary input tokens, cache write tokens, cache read tokens, and output tokens. Multiply each measured category by its applicable rate, then add the results. Do not estimate cost from character count, request count, or the visible size of a prompt alone.

Quanta meters calls on real tokens and reports cache writes and cache reads separately. That gives teams a clearer view of whether a large workload is dominated by new input, initial cache creation, repeated reuse, or generated output. The same accounting method applies when the workload moves between supported models such as Gemini 3.1 Pro and Claude Sonnet 5.5, although the applicable rates and cache behavior can differ.

06 · Step 6: Compare reuse against cache churn

Inspect how often the stable prefix changes. A small edit to shared instructions, a rotating timestamp, injected request identifier, or frequently changing tool definition can create a new cache key. The result may be many writes with few successful reads.

Track cache read tokens per write event and per workload. A high read-to-write relationship usually indicates that the prefix is being reused. A low relationship signals that the team should shorten the variable section, stabilize shared content, or reconsider whether caching fits that workflow.

07 · Step 7: Set operational controls before volume rises

Assign budgets by application, model, or environment. Alert on unexpected increases in cache writes, read volume, total input tokens, and output tokens. During a rollout, compare the measured categories before expanding traffic. This is more reliable than assuming that a cache feature will reduce every bill.

Quanta fits this workflow when a team wants one pay-as-you-go balance across supported models and tools, with usage measured on real tokens. Teams can open a Quanta account to inspect cache writes and reads alongside the rest of their API usage. Referral partners can also earn 19% of referred top-ups, but that does not change the underlying token accounting.