Use Cases + Implementation

How do you reduce the cost of research agents with model routing?

Research agents read many sources and iterate, so cost grows with breadth and depth. Route per-source summarisation and extraction to cheap tiers, reserve strong models for synthesis, contradiction resolution and final reporting, and enforce stopping rules — source caps, time budgets and diminishing-returns checks — so a question cannot consume unbounded budget.

Key facts

Router modelplugsky-fusion escalates per step across tiers (live)
AgentsMemory, tools and orchestration are live
Models30+ models; per-source reading on cheap tiers, synthesis escalated
Loop controlsSource caps, time budgets and stopping rules in your orchestrator
Long contextplugsky-longctx for final synthesis across many summaries
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
RoadmapThe batch endpoint is coming soon for offline reading jobs

TL;DR

  • Read widely on cheap tiers; synthesise narrowly on strong ones.
  • Cap the number of sources per task and per sub-question.
  • Compress sources into structured summaries before synthesis.
  • Stop on diminishing returns, not when the budget empties.
  • Measure cost per brief alongside citation quality.

How it works, step by step

  1. Decompose the research question into sub-questions with explicit source budgets.
  2. Retrieve candidate sources and deduplicate before reading anything.
  3. Summarise each source on a cheap tier into a fixed schema: claims, evidence, confidence, source id.
  4. Detect contradictions and gaps cheaply, then escalate targeted follow-up reads.
  5. Synthesise the brief on a strong tier using compressed summaries rather than raw documents.
  6. Enforce stopping rules: source cap, time budget and a relevance threshold for new material.
  7. Track cost per brief, citation accuracy and reviewer corrections.
1Decompose theresearch questioninto sub-questions2Retrieve candidatesources anddeduplicate before3Summarise eachsource on a cheaptier into a fixed4Detectcontradictions andgaps cheaply, then5Synthesise thebrief on a strongtier using6Enforce stoppingrules: source cap,time budget and a

Try it yourself

Open the AI agent cost calculator →

Reading loops dominate

A research agent's cost is proportional to how much it reads and how many times it re-reads. Each source consumes input tokens, each summary joins the working memory, and iterations resend accumulated context. A thorough question can touch dozens of documents before synthesis begins.

That makes cheap-tier per-source processing the highest-leverage decision. Reading, extracting and summarising do not need the strongest model; they need consistent schemas and reliable compression. Save the expensive model for the step where judgement actually changes the output: synthesis.

Summarise cheap, synthesise strong

Give each source a structured summary — claims, evidence, confidence, source identifier — produced on a cheap tier with a tight schema. Synthesis then works from summaries rather than raw documents, which shrinks context and improves traceability.

  • Deduplicate sources before reading; syndicated copies waste budget.
  • Detect contradictions with a cheap comparison pass and escalate only those pairs.
  • Cap follow-up reads per sub-question to prevent depth spirals.
  • Use plugsky-longctx deliberately for final synthesis across many summaries.

Stopping rules and measurement

Agents that never stop are the most expensive agents. Define stopping rules before the run: a maximum number of sources, a wall-clock budget, a relevance threshold below which new sources are ignored, and a diminishing-returns check that compares new claims against the existing summary set. When a rule fires, produce the brief with what is known and state the gaps.

Measure cost per brief, citation accuracy and human correction rate on a set of realistic questions. A cheap brief that needs heavy rework is not cheap. Start on the free plan with plugsky-micro and plugsky-lite for the reading stage, then evaluate strong and long-context models for synthesis during the 14-day full-access trial; the live pricing page lists plans.

Honest comparison

Research agent choiceRole-routed research agentStrong model reads everythingNo budget discipline
Per-source readingCheap tier with schemasFrontier price per sourceUnbounded sources
SynthesisStrong tier via compressed summariesNative strengthUnbounded context
ContradictionsCheap detection, targeted escalationAll pairs compared stronglyMissed
StoppingSource, time and relevance rulesToken limit onlyNone
EvidenceCost per brief plus citationsToken totalsNone

Frequently asked questions

Why are research agents expensive?

They read many sources and iterate. Each source consumes input tokens, each summary joins working memory, and iterations resend context, so cost scales with breadth and depth rather than with the final brief.

Which model should read sources?

A cheap tier with a structured summary schema. Reading and extraction are consistent, checkable work; save strong models for synthesis where judgement changes the outcome.

How do I stop an agent reading forever?

Set source caps per sub-question, a wall-clock budget, and a diminishing-returns rule that ignores sources adding no new claims. Finish with what is known and state gaps.

Should synthesis use raw documents?

No. Work from structured summaries with source identifiers, which keeps context small and traceable. Use a long-context model deliberately when many summaries must be weighed together.

How do I handle contradictions?

Detect conflicting claims with a cheap comparison pass, then escalate only those pairs to a stronger model rather than comparing every source against every other.

Is batch reading available?

Not yet — the batch endpoint is coming soon. Today, run reading stages with bounded concurrency and a durable queue, deduplicating sources first.

How do I measure research quality?

Cost per brief alongside citation accuracy and reviewer correction rate on realistic questions. Cheap briefs that need rework are a false economy.

Can I prototype for free?

Yes. plugsky-micro and plugsky-lite are on the free plan with no card, and the 14-day full-access trial covers strong and long-context models.