The Playbook
The operator playbook for a credit-metered frontier model, generalized beyond any one vendor. The premise: frontier models are becoming metered utilities, so the operator skill is knowing which tasks deserve frontier pricing and routing everything else to models that are already good enough.
The five moves
Pay frontier prices for judgment, not keystrokes. The five moves as a single operator loop:
1. Reserve the frontier model for long-horizon work only. Its edge is autonomy over hours, not intelligence per prompt. One-session, one-file tasks never touch it.
2. Run the orchestrator pattern. The expensive model touches the project twice — plan, then review — while cheap/local models do the keystrokes:
Commit ready plans under .plans/ and start executors with /work (or orchestrate.py --plan-file) so handoff is file-based, not chat archaeology. Set a project Preferred orchestrator (anchor --set-orchestrator …); if unset, a frontier/near-frontier session may act as temporary coordinator (inventory plans, propose Depends on). For always-on hardware at several skill levels, use /fleet-watch so each tier only claims fit-appropriate, dependency-ready plans—see Fleet workers.
3. Tune prompts on a cheap model first. A sloppy prompt costs the same as a great one. Have a cheap model rewrite every task into a spec with acceptance criteria, files in scope, and a definition of done. Three attempts at a task is the silent budget killer; one tuned attempt is the fix. (scripts/prompt_tuner.py)
4. Don't pay the classifier tax. Security-adjacent work may get rerouted by safety classifiers anyway — route it yourself to the model you'd be rerouted to, and save the credits.
5. Benchmark your real workload. Don't take routing tables on faith — run your own tasks across your own tiers and let pass-rate and latency decide. (scripts/benchmark.py)
The Savings sketches show how large that gap can get — please consider donating to help support this project.
Why this matters double for Anchor
The playbook's economics assume "cheap model" means Sonnet. Anchor pushes it further down: the same orchestration discipline lets a swarm of cheap, always-on workers do the keystrokes. And the discipline that saves money on frontier credits is the same discipline that makes small models reliable at all — small models don't fail because they lack knowledge for scoped tasks; they fail because nothing imposes process on them. Impose it, verify externally, and an 8B model executing a well-cut task spec is indistinguishable from a much bigger model on most of your backlog.
For a typical build, ~80% of the work never needed the big model. Anchor exists to make that 80% run on hardware you own:
Turn that mix into day / month / year dollar sketches (solo → team → org) on the Savings page — with the unit model spelled out so you can plug in your own token rates. If those numbers look familiar, please consider donating to help support this project.