Shift Log: The Cheaper Token Rate Lost at Checkout

The invoice evidence comes before the rate-card verdict

The duty-desk question is simple: which route actually costs less for the workload we ran? A lower advertised input rate does not settle it. The answer lives in the invoice quantities, the applicable rates, and any charges outside token usage.

No real invoice was supplied for this article, so the figures below are an explicitly hypothetical reconciliation worksheet, not a customer bill or a statement of any provider’s current prices. Use the same procedure against an actual invoice, its usage export, and the rate card effective during that billing period.

Our head-to-head compares two fictional routes serving the same model and workload. If you are evaluating Claude Haiku 4.5, for example, keep that exact model constant. Changing models turns a billing comparison into a quality-and-cost evaluation, which needs separate testing.

The first three lines expose the actual token mix

The sample workload contains eight million input tokens and one million output tokens. Of those input tokens, two million qualify for a separately priced cached-read category. That leaves six million ordinary input tokens. The cached tokens are a subset of input, not an extra two million tokens to add on top.

• Route A charges an illustrative $1.00 per million ordinary input tokens, $0.10 per million cached-read tokens, and $4.00 per million output tokens.
• Route B charges an illustrative $0.80 per million ordinary input tokens, $0.40 per million cached-read tokens, and $5.00 per million output tokens.
• These invented rates exist only to demonstrate the arithmetic and are not attributed to Relay365 or any model vendor.

For each invoice category, calculate cost as billed tokens divided by 1,000,000, multiplied by that category’s rate. Confirm whether exported quantities are raw tokens or already expressed in millions before multiplying.

Route A wins the usage subtotal by forty cents

• The ordinary-input line costs Route A $6.00, calculated as six million tokens at $1.00 per million.
• The cached-read line costs Route A $0.20, calculated as two million tokens at $0.10 per million.
• The output line costs Route A $4.00, bringing its usage subtotal to $10.20.
• The same three lines cost Route B $4.80, $0.80, and $5.00 respectively, bringing its usage subtotal to $10.60.

Route B advertises the cheaper ordinary-input rate, but Route A wins this workload. Its savings on cached reads and output exceed its higher ordinary-input charge. That is the first verdict: compare the weighted bill, not the smallest rate-card number.

A cache miss reverses the winning route

Now rerun the comparison with the same eight million input tokens and one million output tokens, but no input qualifying for cached-read pricing. Route A costs $8.00 for input plus $4.00 for output, totaling $12.00. Route B costs $6.40 plus $5.00, totaling $11.40.

Route B now wins by $0.60. Treat cache eligibility as measured behavior, not a permanent discount. Check the actual billing categories: cache writes, cache reads, and ordinary input may have different rules. Only count a discount where the usage record supports it.

The amount due can overturn the usage winner again

Return to the cached workload. If Route A also has an illustrative $1.00 service fee and Route B has none, their pre-tax totals become $11.20 and $10.60. Route B wins checkout despite losing the token subtotal.

Reconcile credits, taxes, currency conversion, and rounding separately. Match the invoice period to the usage export, and inspect retries or failed requests that appear as billable usage. Do not assume every unsuccessful application request was free, or that a wallet top-up represents consumption during that period.

The handoff verdict favors an auditable checkout total

Choose Route A for the illustrated cached workload without extra fees. Choose Route B for the illustrated uncached workload, or when Route A’s added fee reverses the total. Neither route is universally cheaper.

Relay365 fits teams that want pay-as-you-go access across multiple model families through a relay desk. It works with Claude Code, Codex, Gemini CLI, and OpenAI SDK clients. Create an account if that setup fits your workflow, then run a small representative workload and reconcile its recorded charges before expanding usage.

The final desk rule is to approve the route whose measured workload, billing categories, and final payable amount agree—not the one with the loudest per-million-token headline.