Premium Router

Premium Router v2: Fable brain, Haiku price

Three commands, one principle: Fable does the JUDGMENT, Haiku does the VOLUME. Every routing decision is deterministic and every response carries a cost_log with token counts and estimated USD.

classify (Haiku, cheap, the high-volume mode)

Applies an existing rulebook to one input.

python route.py classify --task-type  --input ""

Flags:

• --premium forces Fable to apply the rulebook (premium execution).
• --verify runs Haiku twice; if the two passes disagree, Fable decides.

distill (Fable, expensive, run once per task)

Has Fable write (or rewrite) the rulebook from your labeled examples. Uses effort high: this cost is paid once and amortized over every request.

python route.py distill --task-type  --examples ""

analyze (Fable judgment over documents, Haiku does the reading at scale)

Give a mission and one or more document files. Routing is automatic:

• corpus under ~6000 tokens: Fable reads it directly at effort low

(orchestration measured NOT to pay on small corpora)

• corpus above the threshold: one Haiku worker per document compresses it

into a brief; Fable reads only the briefs and produces the analysis

python route.py analyze --mission "Prioritize these accounts for this week" \
  --docs thread1.txt thread2.txt transcript3.txt

Flags: --threshold N, --force-workers, --force-direct.

Routing rules (deterministic, coded in route.py)

• Rulebook exists + normal case: Haiku applies it.
• No rulebook: classify refuses with a clear error. Distillation is always an

explicit step, never a side effect. A classify call can never overwrite a saved rulebook.

• Sensitive keyword (legal, procurement, enterprise contract, security) or

--premium: Fable applies the same rulebook.

• --verify disagreement between two Haiku passes: Fable decides.
• analyze routes on measured corpus size, never on model self-assessment.
• Never escalate on a model-declared "confidence" score.

Measured anti-patterns (do NOT do these)

• Advisor Tool (Haiku executor + Fable advisor): measured ~19x MORE expensive

than Fable alone. It boosts quality, it does not save money.

• Orchestrating small corpora: briefs barely compress short documents; below

~1.7x compression the workers cost more than they save.

• Prompt caching short rulebooks: ignored below ~4096 tokens on Haiku 4.5.

Requirements

• ANTHROPIC_API_KEY in the environment.
• The anthropic package installed (see requirements.txt).
• An org that is NOT on zero-data-retention (Fable returns 400 otherwise).

Pricing context

Fable 5: $10 / $50 per million tokens (in / out). Haiku 4.5: $1 / $5. Fable at effort low often exceeds the max performance of previous frontier models (per Anthropic docs), so effort low is the default everywhere except distillation.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: ToolMonsters
• Source: ToolMonsters/premium-router
• License: MIT

Install and usage instructions live in the source repository linked above.