$8.09 to Triage 10,000 Emails: A Planning Receipt

01 / The $8.09 receipt starts with assumptions, not promises

For 10,000 incoming emails, a small-model-first triage pipeline can have an illustrative token bill of $8.09. That is a planning calculation, not a measured production result or a current price quote. The useful part is the receipt: you can replace each assumption with your own rates, token counts, and escalation percentage.

Consider DeepSeek V4 Flash for the first pass and Claude Opus 5 for difficult messages. For this example only, assign the small tier budget rates of $0.20 per million input tokens and $0.80 per million output tokens. Assign the large tier $3.00 input and $15.00 output. These are hypothetical worksheet rates, not advertised prices for either model. Check current rates before deployment.

02 / Give the small model one narrow sorting job

The first pass classifies each email into billing, account access, sales, routine support, or other. It also returns urgency, a short reason, and an escalation flag. It does not send replies, approve refunds, or change account settings.

Budget an average of 800 input tokens and 80 output tokens per first-pass request. Input includes instructions, the relevant email text, and necessary metadata. Trim repeated signatures and quoted threads, but preserve attachments or earlier context when they affect the decision. Those averages are assumptions to verify on representative messages.

• ■ 10,000 emails × 800 input tokens = 8,000,000 input tokens.
• ■ 10,000 emails × 80 output tokens = 800,000 output tokens.
• ■ Input charge: 8 × $0.20 = $1.60.
• ■ Output charge: 0.8 × $0.80 = $0.64.
• ■ First-pass subtotal: $2.24.

03 / Escalate 1,000 emails without trusting confidence alone

Assume 10% of messages need the large model. That is a budget scenario, not an accuracy claim. Escalate messages with conflicting categories, missing context, suspicious instructions, or sensitive billing disputes. A confidence score can be one signal, but a model calling itself certain is not evidence that it is correct.

Give each fallback request 1,200 input tokens, including fuller context and the first-pass result, and allow an average of 150 output tokens. Ask the large model to reassess the original email rather than merely endorse the first classification.

• ■ 1,000 escalations × 1,200 input tokens = 1,200,000 input tokens.
• ■ 1,000 escalations × 150 output tokens = 150,000 output tokens.
• ■ Input charge: 1.2 × $3.00 = $3.60.
• ■ Output charge: 0.15 × $15.00 = $2.25.
• ■ Fallback subtotal: $5.85. Combined token bill: $8.09.

04 / The comparison receipt is $58.50, with conditions

Sending all 10,000 emails directly to the large tier, using the same 1,200-input and 150-output token allowance, would cost $36.00 for input plus $22.50 for output: $58.50. The staged scenario therefore saves $50.41, about 86%, under these assumptions. This compares two workflow budgets; it does not establish equal classification quality.

Fallback frequency drives the result. At the assumed token lengths and rates, each escalation costs $0.00585. Every additional 1,000 escalations adds $5.85. At 20% escalation, the total becomes $13.94; at 50%, $31.49. Retries, longer threads, and malformed responses add costs too.

05 / Keep the production bill separate from the worksheet

CheapRelay provides pay-as-you-go access to both tiers through one key and OpenAI-compatible routes under /v1 at https://api.cheapapirelay.com. Billing follows real tokens, including cache reads and writes, with no monthly fee or subscription. This receipt assumes no caching and excludes storage, attachment extraction, human review, and engineering time. Sign up for CheapRelay to test the workflow with a capped budget and current model rates.

Before scaling, label a representative sample and review missed urgent emails, unnecessary escalations, and category errors. Treat email content as untrusted data, validate the response schema, and keep irreversible actions behind separate authorization. Log actual token usage and escalation reasons. A cheap receipt is useful only when the sorting remains dependable.