Three composable levers that cut the token bill
Amplifier agents run on frontier models billed per token. The foundation instruction surface alone measures 27,025 est. tokens; add agent prompts and history and a turn easily carries 40k+ tokens — re-sent every single turn.
The first lever attacks exactly this repetition — and it's already on.
On Sonnet 4.5, cached input reads at $0.30/M versus $3.00/M fresh — exactly 10%. In the Anthropic provider, enable_prompt_caching defaults to True, so the durable savings lever is live with zero configuration.
Free by default — but naive usage still leaves money on the table.
Without caching applied, that 40k+ context is billed at full $3.00/M every turn. And sending simple parsing or classification work to a frontier model means paying top rates for tasks a cheap model handles fine.
Two more levers close that gap — here's how each one works.
The provider stamps cache_control on the system prompt, tools, and the last message. Re-sending a 40k context over 10 turns drops from ~$1.20 to ~$0.258 on input alone — about 78.5% saved on the repeated context.
That's the default lever. The second one is a single declaration away.
~78.5% saved on repeated context (illustrative model built from verified Sonnet 4.5 rates, not a measured bill).
Agents declare a semantic model_role. The 'fast' role — parsing, classification, file ops, bulk work — routes to Haiku instead of Opus, which is 5x cheaper per token on both input and output.
The third lever governs what happens once the loop is running.
Amplifier's streaming loop executes multiple tool calls concurrently with deterministic ordering. A reference cost-aware scheduler hook watches tool selection and can veto or swap an expensive choice against a cost budget.
Stack all three, and the bill changes shape entirely.
Caching alone saves ~78.5% on repeated context; routing cuts trivial work 5x; the loop bounds and governs the rest. Composed, the optimizations are order-of-magnitude — and the biggest lever, caching, needs zero configuration.
The ~$50 → ~$5 figure is an illustrative headline, not a measured bill — the defensible numbers are 10%, ~78.5%, and 5x.
Turn on nothing to get caching. Declare a model_role to route cheap work to cheap models. Let the loop govern the rest. Stack cheap-by-default levers rather than hand-tuning a bill.
The order-of-magnitude win comes from composition, not configuration.
Data as of: July 21, 2026. Grounded in Amplifier repos under /home/ramparte/dev/ANext plus the primary-source cost-aware scheduler.
Feature status: Caching SHIPPED & on by default; routing-matrix SHIPPED; loop-streaming SHIPPED; cost-aware scheduler is a reference implementation.
Commands run:
grep -n 'enable_prompt_caching' amplifier_module_provider_anthropic/__init__.py — defaults True (line 590); cache_control on system/tools/last message.sed -n '43,60p' amplifier_module_provider_anthropic/_cost.py & python3 — Sonnet 4.5 input $3.00/M vs cache_read $0.30/M (exactly 10%).python3 illustrative model — 10 turns of 40k: ~$1.20 naive vs ~$0.258 cached = ~78.5% (from verified _RATES, not a measured bill).cat amplifier-foundation-budget-gate/budgets.json — root surface measured 27,025 est. tokens on 2026-02-10.cat routing-instructions.md & sed -n '43,60p' economy.yaml — 13 roles incl. 'fast'; Opus $5/$25 vs Haiku 4.5 $1/$5 = 5x.sed README.md (loop-streaming) & git log (recipes dfe6875) — concurrent tool calls, deterministic order, parallel cap 3.read /tmp/cost-aware-repo/__init__.py (208 lines) — hook on tool:selecting (priority 60); continue/modify/deny against max_cost.Gaps: The ~$50 → ~$5 headline is a rounded, illustrative figure — NOT found verbatim in any repo. The 40k+/turn figure is inferred (measured value is the 27,025-token root surface). Scheduler tool_costs are example units, not USD.
Primary contributors: Brian Krabach (sole author of all 14 scheduler commits; top contributor with 38 commits to routing-matrix).