Use Cases + Implementation

How do you reduce the cost of AI cost optimization with model routing?

AI cost optimization with model routing means matching each request to the cheapest model that can handle it. Audit where tokens actually go, route routine classification and drafting to plugsky-micro or plugsky-lite, escalate ambiguous or high-value requests to stronger tiers, and cap context, output length and retries. Measure routing accuracy on an evaluation set before widening the rules.

Key facts

Router modelplugsky-fusion fans out across tiers and escalates per request (live)
Strategiescost_saver, balanced, max_quality and custom rules
Models30+ models from free to frontier behind one endpoint
PricingFlat monthly self-serve plans with no per-token charges on self-serve
Free tierplugsky-micro and plugsky-lite on the free plan, no card required
Cost controlsScoped keys and usage analytics per key and workspace
RoadmapClassifier routing (/v1/plugsky/route, model=auto) is coming soon
DeploymentCloud, VPC, on-prem and air-gapped options

TL;DR

  • Audit token spend by feature before changing any model.
  • Route the bulk of routine calls to cheap tiers and escalate the rest.
  • Cap context, output length and retries — they usually explain cost spikes.
  • Use fusion strategies so routing is configuration, not code.
  • Track cost per resolved task, not cost per token.

How it works, step by step

  1. Instrument usage by feature, model and key to find where spend actually concentrates.
  2. Classify requests by complexity using length, intent and required accuracy.
  3. Set a default cheap tier per workload and an escalation rule for hard requests.
  4. Apply cost_saver, balanced or max_quality strategies per workspace or key.
  5. Trim prompts and retrieved context, and cap max_tokens per task type.
  6. Evaluate quality on a fixed task set at each tier before removing the strong model.
  7. Review routing logs monthly and adjust thresholds as prompts and traffic change.
1Instrument usage byfeature, model andkey to find where2Classify requestsby complexity usinglength, intent and3Set a default cheaptier per workloadand an escalation4Apply cost_saver,balanced ormax_quality5Trim prompts andretrieved context,and cap max_tokens6Evaluate quality ona fixed task set ateach tier before

Try it yourself

Open the LLM API cost calculator →

Where AI budgets actually leak

Most overspend comes from three habits: sending every request to the strongest model, resending context that has not changed, and letting retries and agent loops run unbounded. A routing review starts with measurement — tokens and requests by feature, model and key — because the workload that dominates the bill is rarely the one teams expect.

Once you know the shape, tier the traffic. High-volume, low-ambiguity work such as classification, extraction and short replies tolerates cheap models well. Complex reasoning, long synthesis and anything customer-visible at high stakes deserves stronger tiers. Routing turns that judgement into configuration rather than per-call engineering.

Routing as configuration, not code

Plugsky's fusion router runs a default chain and escalates when a request is harder than the tier handles. Strategies make the policy explicit:

  • cost_saver for bulk, tolerance for occasional retries.
  • balanced for interactive features where quality and spend both matter.
  • max_quality for work a person will act on.
  • Custom rules for known high-value paths that must never be downgraded.

Set the strategy per workspace or key so a batch pipeline and a live product do not share one policy. Classifier routing with model=auto is coming soon; until then, fusion strategies and per-endpoint aliases cover most cases.

Measure savings without losing quality

The right metric is cost per resolved task, not cost per token. Build a fixed evaluation set per workload, run it at each tier, and compare task success, retries and escalation behaviour. A cheaper tier that doubles retries is not cheaper, and a strong tier reserved for the hardest requests usually beats both extremes.

Guardrails matter too: cap context size, cap output length, deduplicate prompts and cache stable prefixes where the API allows. Then review monthly — prompts, traffic mixes and model catalogues all drift. Flat-rate self-serve plans remove per-token billing from the equation; see the live pricing page for plan details, and start on the free plan with plugsky-micro and plugsky-lite before scaling.

Honest comparison

Cost leverPlugsky model routingAlways-on frontier modelOne cheap model everywhere
Routine requestsCheap tiers by defaultFrontier price per callCheap, sometimes too weak
Hard requestsAutomatic escalationNative strengthNo headroom
Policy controlStrategies and custom rulesNoneNone
RetriesNo per-token charge on self-serve flat plansRetries add costRetries add cost
Quality evidencePer-request model logsUniform but expensiveUniform but limited

Frequently asked questions

What is model routing in AI cost optimization?

It sends each request to the cheapest model that can handle it, escalating to stronger tiers only when needed. Routing policies live in configuration, and the router logs which model answered.

How do I know which requests are cheap to serve?

Instrument usage by feature and look at length, ambiguity and required accuracy. Classification, extraction and short replies are usually safe for cheap tiers; reasoning and synthesis are not.

Does routing hurt quality?

It can if thresholds are wrong. Evaluate each workload on a fixed task set at every tier, watch retries and escalation rates, and keep high-stakes paths on max_quality until evidence says otherwise.

What are fusion strategies?

cost_saver, balanced and max_quality preset routing behaviour, with custom rules available. Set them per workspace or key so pipelines and live products do not share one policy.

Is there a classifier router?

Not yet — classifier routing with model=auto is coming soon. Today, fusion strategies and per-endpoint aliases handle most routing needs.

How does flat pricing change the maths?

On self-serve plans there are no per-token charges, so retries and prompt growth do not change the bill the way they do on metered APIs. Routing still matters for latency, throughput and fair-use headroom.

What should I measure to prove savings?

Cost per resolved task alongside success rate, retry count and escalation rate. A tier that looks cheap but needs more attempts is not saving anything.

Can I try routing for free?

Yes. plugsky-micro and plugsky-lite are on the free plan with no card, and a 14-day full-access trial covers stronger tiers for evaluation.