$5.20 to Triage 10,000 Emails: A Two-Model Receipt

01 / The $5.20 receipt starts with explicit assumptions

Run every email through a small model, then send uncertain cases to a bigger one. The economics are straightforward; the routing policy is the work. Here is a planning example for 10,000 emails, with a total model bill of $5.20.

These are illustrative rates, not current CheapRelay prices or a measured production result. Assume the small model costs $0.20 per million input tokens and $0.40 per million output tokens. Assume the bigger model costs $2.00 per million input tokens and $8.00 per million output tokens. Replace those four figures with live rates before budgeting.

Candidate roles could be DeepSeek V4 Flash for the first pass and Gemini 3.1 Pro for escalation. Test both on your own mailbox; their names alone do not establish accuracy or suitability.

02 / Give each email a 1,000-token input budget

The job is classification, not correspondence. Each request receives a normalized email excerpt, sender metadata, a short label policy, and instructions to return structured fields: category, priority, suggested queue, and escalation reason.

Budget an average of 1,000 input tokens and 100 output tokens per request. These are workload assumptions, not promises about tokenizer behavior. Measure actual usage because long threads, signatures, and verbose answers can move the bill.

The bigger model gets its own self-contained payload within the same average budget, not an appended transcript of every previous decision. For an apples-to-apples cost comparison, use that same budget when calculating the all-big-model baseline.

03 / Receipt: 10,000 first passes and 1,000 escalations

• ■ Small-model input: 10,000 × 1,000 = 10 million tokens. At $0.20 per million: $2.00.
• ■ Small-model output: 10,000 × 100 = 1 million tokens. At $0.40 per million: $0.40.
• ■ Small-model subtotal: $2.40.
• ■ Big-model input: 1,000 × 1,000 = 1 million tokens. At $2.00 per million: $2.00.
• ■ Big-model output: 1,000 × 100 = 100,000 tokens. At $8.00 per million: $0.80.
• ■ Big-model subtotal: $2.80.
• ■ Total: 11,000 requests and $5.20, or $0.00052 per incoming email.

Sending all 10,000 emails directly to the bigger model would cost $20.00 for input and $8.00 for output: $28.00 total. Under these assumptions, the two-stage route saves $22.80, approximately 81.4%. That is arithmetic, not an accuracy benchmark.

04 / A 10% fallback rate is a target, not evidence

Do not let a model's self-reported confidence be the only gate. Escalate missing fields, invalid categories, contradictory instructions, ambiguous ownership, and cases involving sensitive account or payment changes. Keep a human-review queue for decisions that could cause harm.

Treat email text as untrusted data. Instructions inside a message must not override the triage policy. Neither model should send replies, delete messages, execute attachments, or authorize payments in this workflow.

Evaluate routing against a manually labeled sample. Count wrong urgent labels and missed urgent messages separately. A cheap route that silently buries important mail is not a saving. Stronger models can also be wrong, so escalation is not a correctness guarantee.

05 / The fallback rate controls the next receipt

With these token budgets, each bigger-model call costs $0.0028. At 20% fallback, the total becomes $2.40 plus $5.60, or $8.00. At 50%, it becomes $16.40. At 100%, it reaches $30.40: more than the $28.00 baseline because every email pays for both passes.

CheapRelay provides pay-as-you-go access through one key and an OpenAI-compatible API, with no monthly fee or subscription. Sign up for CheapRelay to run a small, labeled pilot before committing your mailbox. Billing follows real tokens, including cache reads and writes; this receipt assumes no caching and excludes retries, hosting, and human review.

Log tokens, escalations, retries, and reviewed errors. Rebuild the receipt from actual usage, then decide whether the cheaper route earns its place.