Model Routing and Cost Control (Delivery)

Model Routing and Cost Control (Delivery)

1. 01

   Inventory every recurring client deliverable and tag it by reasoning depth

   Split work into three buckets: single-pass drafting, multi-step analysis, and agentic workflows that touch repositories, databases, or external APIs. OpenAI's October 2, 2026 GPT-6 guide separates model tiers along roughly these same lines, so the inventory maps cleanly onto tier selection.

2. 02

   Assign a default assistant per bucket rather than per person

   Standardise one primary assistant for copy and research, one for code and file-heavy analysis, and hold a third slot for agentic runs. Rotating between Claude, ChatGPT, and Anthropic API access per individual preference destroys cost visibility and makes retainer pricing guesswork.

3. 03

   Run a paid pilot on the two highest-volume deliverable types before committing seats

   Push ten real client jobs through each candidate tier and record wall-clock time, revision rounds, and token or seat cost. A September 3, 2026 technical analysis found that top benchmark scores do not reliably predict production performance, so pilot output beats leaderboard position.

4. 04

   Set a per-deliverable cost ceiling and a monthly seat budget before the retainer is quoted

   Convert pilot data into a cost-per-deliverable figure, then multiply by expected monthly volume. Serverless access to frontier open models, such as Prime Intellect's October 2, 2026 launch, gives agencies a cheaper overflow path when client workloads spike.

5. 05

   Route agentic tasks through a permission review before they touch client systems

   Apple's October 2026 tightening of macOS Full Disk Access exists because capable agents reading files, mail, and browsing history widen the blast radius. Confirm which local folders and connected apps each agent can reach, and document the answer in the client file.

6. 06

   Log model version, tier, and date against every deliverable that ships

   Model families change monthly, so a deliverable produced in October may not be reproducible in December. Version logging lets you answer client questions with tested evidence instead of vendor summaries.

7. 07

   Review routing quarterly against actual invoice and retention data

   Pull cost per deliverable, revision rate, and client churn for the quarter, then re-tier any bucket where cost rose without a matching quality gain. Routing decisions that never get re-tested quietly become the margin leak.