Short Prompts Save Money When Context Stops Growing

Splitting a large assignment into smaller prompts feels economical: each request looks leaner, and each answer is easier to inspect. Yet the invoice measures tokens across the entire workflow, not the size of its largest prompt. Repeated instructions, accumulated answers, and retries can make the smaller requests more expensive.

The useful comparison is therefore not prompt length alone. It is the cost of producing an acceptable result, including every request needed to get there.

Step one fixes the deliverable before counting tokens

Consider a documentation task with four independent sections. The source material contains 12,000 tokens, divided evenly between the sections. Shared instructions add another 1,000 tokens, and the finished document should contain 2,000 output tokens.

Compare two designs: one request handles all four sections, or four requests each handle one section. Keep the model, quality requirements, and total target output identical. Otherwise, a cheaper result might simply be a smaller or weaker deliverable.

Step two prices the complete single request

Use illustrative rates of $2 per million ordinary input tokens and $10 per million output tokens. These are arithmetic assumptions, not a quoted tariff for any model.

The single request contains 13,000 input tokens: 12,000 source tokens plus 1,000 instruction tokens. Its input costs $0.026. Generating 2,000 output tokens costs $0.020, giving a total of $0.046.

This baseline assumes the result needs no repair request. It also excludes tool calls and other billable categories that a real workflow might introduce.

Step three charges every repeated instruction block

In the split design, each request receives 3,000 source tokens and the same 1,000 instruction tokens. Across four requests, that becomes 16,000 input tokens. Four answers of 500 tokens still total 2,000 output tokens.

Ordinary input now costs $0.032, while output remains $0.020. The complete split workflow costs $0.052, approximately 13 percent more than the single request. Smaller prompts have increased spending because the shared instructions were purchased four times.

If later requests also receive earlier answers, the bill grows again. Forwarding all previous section outputs adds 500, 1,000, and 1,500 input tokens to the subsequent requests. Those additional 3,000 tokens raise the total to $0.058.

Step four applies caching without assuming free context

Now suppose the provider supports caching the shared instruction prefix. Assume a cache write costs $2.50 per million tokens and a cache read costs $0.20 per million tokens. Keep ordinary input and output rates unchanged.

For four independent requests, the 12,000 source tokens cost $0.024. Writing the 1,000-token prefix once costs $0.0025. Reading it three times costs $0.0006. Add $0.020 for output, and the cached split workflow totals $0.0471.

Caching nearly closes the gap, but does not reverse it here. This calculation assumes the prefix qualifies, remains reusable, and produces three actual cache hits. Minimum cache sizes, expiration, and provider rules can invalidate that assumption.

Step five measures successful delivery rather than attempts

Retries can change the winner. Repairing one flawed section may be cheaper than regenerating an entire document, provided the repair receives only the context it needs. Conversely, a final consolidation request can erase those savings.

Run representative tasks and record total charges, acceptance rate, and elapsed time. Divide total spending by accepted deliverables. Track latency separately: parallel requests may finish sooner without costing less.

Step six reconciles estimates with provider billing

BestRelay publishes this walkthrough as a quality-first API relay forwarding each request to the official model named in it, without substitution. Its coverage includes Claude, GPT, Gemini, DeepSeek, Kimi, GLM, Qwen, and Grok. It works with Claude Code, Codex, Gemini CLI, and any OpenAI SDK.

Its pay-as-you-go billing uses provider-reported token counts, with cache reads and writes priced separately. Reconcile those categories against your worksheet rather than estimating from visible text alone. Members also earn referral commission when invited users top up; keep that separate from workload costs.

Create a BestRelay account to test your own workload. Choose the workflow with the lowest measured cost per accepted result, not the shortest-looking prompt.