Price the Whole Email Brief Before Crossing a Prompt Tier

Affiliation: Marketing Wiki's commissioning editor maintains Migma. Keep the approved facts for an email in Migma, then budget any separate model call against the complete prompt it will receive. A lower advertised token price does not tell you the cost of a brief that crosses a prompt-length tier.

Marketing Wiki Research Automation published this guide directly without independent review. Sources were checked on October 7, 2026. The calculations below are synthetic examples using published prices, not measured bills or Migma charges.

Start with the brief that must survive#

Migma's brand setup guide describes Brand Guidelines for approved information and AI Instructions for lasting preferences. Use those surfaces to hold reviewed positioning and constraints for subsequent email work. A separate assistant can help extract a larger source packet, but its output still needs approval before it becomes the campaign brief.

Suppose a training company has course syllabuses, refund cutoffs, accessibility arrangements and archived workshop announcements. The current workshop's refund deadline must survive extraction. An outdated campaign example may be removable. Those two kinds of context should not compete merely because both occupy tokens.

This guide concerns a separate API budgeting decision. The cited Migma documentation does not establish a native Haiku selector, shared billing or automatic management of Anthropic's context tiers.

Read the length condition beside the price#

Anthropic's October 7 Haiku 5.5 announcement publishes different rates for prompts up to and over 100,000 tokens. Its table lists these input and output rates in US dollars per million tokens:

The condition is prompt length, not the number of words in the marketer's visible instruction. Count the full submitted context with the supported model tooling, including material added by the application. The announcement also says the model has an updated tokenizer. Do not reuse an old model's count as a confirmed count for this one.

For a request without cached input, the worksheet is:

Select the published tier using complete prompt length.
Input charge = input tokens / 1,000,000 × tier input rate
Output charge = output tokens / 1,000,000 × tier output rate
Illustrative request charge = input charge + output charge

An assumed 80,000-input-token request with 1,000 output tokens gives $0.008 + $0.0005 = $0.0085. An assumed 120,000-input-token request with the same output gives $0.06 + $0.0025 = $0.0625. These examples apply each listed rate to its respective request. They are not marginal-band calculations that price only the last 20,000 input tokens differently.

Actual usage, caching, retries and account terms can change the bill. The launch table also lists separate cache read and write prices; this simple worksheet deliberately assumes no cache. Use the current provider accounting rules for a real request rather than treating the illustration as an invoice.

Remove duplication before removing safeguards#

Compare three candidate packets before choosing one:

A smaller packet is useful only if it still supports the assignment. If the task needs several independent sources, split extraction by source and preserve identifiers in the handoff. Then check whether combining the results loses conflicts or exceptions. Multiple calls can also add cost; do not label splitting a saving before counting all attempts.

Approve the budget and the facts separately#

Record the model identifier, pricing access date, complete input count, assumed output count, cache assumption and number of permitted attempts. Set a pause point for another extraction pass. Budget approval permits the bounded computation; it does not approve the email's claims or a send.

Move only the accepted fact card into the Migma brief. Keep a reference to the source packet so a reviewer can inspect a restriction that was shortened. After the run, reconcile provider usage using the separate estimate-to-actual ledger. Begin with one representative brief, count it using the intended model, and compare two evidence-preserving packets before choosing a tier.