LLM Council Cost

LLM Council: Cost

Purpose

Cut the cost and latency of a multi-model pipeline that already exists — a council, ensemble, debate, or plain "we call N models per request" setup — without giving up the quality it was built for. The lever is almost never "pick a cheaper model everywhere"; it's making the expensive path fire only when it's actually earning its keep.

When to use / when NOT to use

Use this skill when the pipeline is already running and the complaint is cost, latency, or token spend — including when the user never says "council" ("we call 4 models per request and the bill exploded", "cut multi-agent orchestration cost").

Not for:

• Deciding whether to adopt a council at all → sibling llm-council-when. If the user is asking "is this worth it" rather than "make it cheaper", route there instead.
• Choosing how to combine member outputs (voting, weighting, judge aggregation) → sibling llm-council-aggregation.
• Generic single-model cost cutting with no multi-model pipeline in play ("switch to a cheaper model", "our OpenAI bill is high") — that's a model choice, not a routing/gating problem.
• Cloud infra cost (EC2, S3, hosting bill) or database query optimization — unrelated domains that occasionally share the word "cost".

The core move: route first, council second

Put a cheap gate at the front door so the expensive path fires only on the hard tail, not on every request. Escalation ladder, cheapest first:

1. Semantic router — embedding similarity to known intents; escalate below a similarity threshold.
2. Learned router — trained to predict which queries need the strong path. RouteLLM reports ~85% cost reduction while retaining ~95% of GPT-4 quality on MT-Bench (paper-reported, not vendor-reported — see references/routing-and-caching.md#learned-routers). Real operational tax: routers degrade on out-of-distribution traffic and need retraining whenever the model pool changes.
3. Pre-inference screening for Mixture-of-Agents — cheap pre-filtering before the expensive fan-out. RouteMoA reports 89.8% cost / 63.6% latency reduction in large-pool settings (author-reported; see reference).

Real routed traffic tends to skew cheap (roughly 80% cheap tier / 15-20% mid / 5% premium is a commonly observed anecdotal shape, not a law — measure your own split).

Calibration is the trigger, not raw confidence

Never gate escalation on a model's raw self-reported confidence — it is unreliable and can be inversely correlated with correctness on hard tasks. Fit the escalation threshold on your own traffic's calibration curve (ECE/Brier vs ground truth). A verified anchor: isotonic-regression calibration of token-margin uncertainty cut inference cost 31% (95% CI 27-35%) while improving ECE from 0.12 to 0.03 on a 75k-query production NER workload (UCCI; see reference).

For multi-member setups, peer disagreement is a better escalation signal than any single member's self-report: run cheap models first, escalate only when they disagree (2-then-3-judge pattern), rather than always paying for every member.

Know your multipliers before you plan around them

Order-of-magnitude planning numbers only — re-measure on your own workload, don't plan a budget off someone else's paper:

Cost compounds non-linearly on top of these: context re-ingestion across turns, retries, and orchestration overhead can push real overhead well past the naive per-call estimate. Measure cost-per-resolved-outcome, not cost-per-query — a 3x per-query multiplier can still be a net win if it meaningfully cuts a costly downstream error rate, and a net loss if it doesn't.

Caching and batching

Structure council prompts with a shared static prefix (the question plus rubric) first, per-member instructions after, so every member call hits the cache:

• Anthropic prompt caching is opt-in via cache_control: cache writes cost a premium, cached reads are a small fraction of base input price (as of mid-2026 — verify: docs.anthropic.com prompt caching).
• OpenAI caches automatically for prompts ≥1024 tokens, at roughly a 50% discount on the cached prefix (as of mid-2026 — verify: platform.openai.com).
• Batch APIs (~50% discount, as of mid-2026 — verify provider docs) fit offline/eval workloads only. Batching low-traffic interactive calls just adds wait with no cost win on a single request.

Session-aware routing

Don't re-route per turn inside an agent's tool-call loop. Mid-session model switches break prompt-cache locality and can invalidate in-flight tool-call sequences. Lock the model for the duration of an active loop; re-route only at session or task boundaries.

Latency: budget tails, not averages

Fan-out wall-clock time is bounded by the slowest member — one tail-latency member dominates the whole council's latency, regardless of how fast the others respond. Budget P95/P99, not the mean. Use TTFT-based hedging triggers rather than static timeouts, and cap hedge load (roughly a 10% token-bucket) so hedging itself doesn't become a cost problem. Parallel dispatch is a latency lever only — you still pay for every member's tokens regardless of wall-clock time.

Cheap wins (roughly ordered by effort)

• Cap debate rounds with a statistical stopping rule instead of a fixed count — one reported case converged in ~1 round vs a fixed 5, a 3.7x call cut (see reference).
• Drop irrelevant specialist seats per task instead of always dispatching a fixed roster.
• Prune samples past ~5 in self-consistency-style sampling — diminishing returns; temperature matters more than sample count beyond that point.
• Consider Self-MoA: repeatedly sampling your single strongest model can beat mixing weaker ones (+6.6pp AlpacaEval reported; see reference) — and it erases the multi-vendor operational burden entirely.
• Distillation / internalized-debate approaches report matching explicit debate at a fraction of the tokens — directional, single-paper evidence, worth a pilot before a full migration.

Workflow

1. Map the current pipeline's shape — how many models, what topology (self-consistency, MoA, debate, full council), and which multiplier row above it resembles. This tells you the ceiling on savings before you start.
2. Add a route-first gate sized to the traffic: semantic router for a known intent set, a learned router if intents are open-ended, pre-filter screening if the pattern is MoA-shaped.
3. Calibrate the escalation threshold on the user's own traffic — never ship a threshold based on raw self-reported confidence or a borrowed default. For multi-member setups, prefer peer-disagreement escalation.
4. Restructure prompts for caching: shared static prefix first, per-call variation last, so every member/turn hits the cache.
5. Check for mid-session re-routing inside any agentic tool-loop and remove it; move re-routing to session/task boundaries only.
6. Budget latency at P95/P99 and add TTFT-based hedging with a capped hedge budget if tail latency (not cost) is the complaint.
7. Apply the cheap-wins list relevant to the pipeline's topology.
8. Re-measure: cost-per-resolved-outcome before/after, and flag which cited numbers (vendor percentages especially) need re-verification on the user's own workload before being trusted for planning.

Output spec

A complete answer:

• Names the route-first gate design appropriate to the traffic shape.
• States the calibration approach for the escalation threshold (not raw self-report).
• Gives a concrete caching/prompt-structure change if member calls don't already share a cached prefix.
• Flags any mid-session re-routing anti-pattern if an agentic loop is in play.
• Distinguishes cost levers from latency-only levers (parallel dispatch, TTFT hedging) so the user doesn't expect a latency fix to also cut spend.
• Frames savings as cost-per-resolved-outcome, and notes any cited percentage/multiplier should be re-measured on the user's own workload before being used to plan a budget.

Failure modes & gotchas

• Routing on raw self-reported confidence. It's unreliable and can be inversely correlated with correctness on the hard tail — the exact traffic the gate exists to catch.
• Treating a learned router as fire-and-forget. It degrades on out-of-distribution traffic and needs retraining whenever the model pool changes — budget for that as an ongoing cost, not a one-time setup.
• Re-routing mid-session inside a tool-call loop. Breaks cache locality and can invalidate in-flight tool sequences; re-route at boundaries only.
• Conflating latency and cost fixes. Parallel dispatch and hedging help wall-clock time; they do not reduce the tokens paid for. Don't sell a latency fix as a cost fix.
• Trusting a single vendor/paper percentage as a planning number. Every multiplier and discount cited here is order-of-magnitude and version-gated for a reason — re-measure on the actual workload before it goes into a budget.
• Batching interactive traffic. Batch API discounts only pay off for offline/eval workloads; applying them to low-traffic interactive calls just adds wait.

References

• references/routing-and-caching.md — deeper routing, calibration, caching, and latency detail with citations.

Siblings

• llm-council-when — decides whether to adopt a council at all.
• llm-council-architecture — designs the pipeline once a council is chosen.
• llm-council-members — picks council composition and diversity.
• llm-council-aggregation — combines member answers into one output.
• llm-council-prompts — writes stage-specific prompts (debate, synthesis).
• llm-council-failure-modes — defenses against conformity and collapse.
• llm-council-harness — headless CLI implementation of a council.