Route Work by Cost: Models, Mixes, and Metering

Not every task needs the flagship model. A simple routing policy can cut the bill by an order of magnitude.

The default model is the most expensive habit in bot-building. Daily digest bots, log pruners, and format-checkers don’t need the smartest model you have — they need a reliable model. The trick is knowing which is which, and routing deliberately.

The three tiers of bot work

Judgment calls. Strategy, nuanced writing, deciding what matters. Use the strongest model. Wrong here is expensive in rework and trust.

Reliable labor. Summarize, classify, extract, compare-against-list, format. Mid-tier models do this at a fraction of the cost with near-identical quality — the task is bounded, and the prompt patterns do the heavy lifting.

** Plumbing.** “Did anything change? Update the log.” Cheap models, tiny prompts, high frequency. This is where careless builders burn quotas on “nothing today.”

A routing policy in ten lines

When your stack supports model choice (natively, or by connecting other providers), put the policy in the prompt or the coordinator bot:

Use the strong model only for: first drafts I will publish, decisions about money, and anything config.md marks “judgment.” Use the fast model for: summaries, log updates, format checks, and any run whose last output was “nothing to report.” If unsure which applies, ask me once, then remember in config.md.

Two effects, immediately: the bill drops, and the strong model’s attention stops being spent on “is this video new?”

Where the money actually goes

• Frequency × verbosity. An hourly routine with an essay format costs more than a daily one with a strict format. Trim the format before you trim the schedule.
• Re-reading everything. A bot that pulls the full transcript when only new videos matter. State files are a cost tool as much as a memory tool.
• Retry loops. A failing routine that retries hourly burns quota all night. One retry per cycle, then stop and say so — this is also in the troubleshooting loop.
• Quota blindness. Fleet builders keep a usage bot watching the meter and telling the others. Boring, saves weekends.

Connecting cheaper models

A popular pattern (see Connect DeepSeek to your bot): your Grok Bot stays the brain and interface, but delegates bulk work to a cheaper provider’s API via a workspace script. The routing rule above applies unchanged — the tier labels matter more than the brand names.

When NOT to downgrade

• First drafts of anything with your name on it.
• Anything where a subtle misread becomes an irreversible action.
• New use cases, for their first two weeks — you’re still learning what “reliable” means for them.

Downgrade the plumbing, not the judgment. That’s the whole policy.