When should you use a frontier-tier model like plugsky-frontier?

Key facts

Tier	Top capability tier in the Plugsky model ladder
Best for	Hard reasoning, high-stakes drafting and multi-document synthesis
Use as	Escalation target and quality baseline, not a default
Capabilities	Streaming, function calling, JSON mode and long-context
Context class	128K-class window; live limits are published per model
Pricing tier	Paid-plan model; free plan covers plugsky-micro and plugsky-lite
Resilience	Backup upstream plus same-profile fallback peers
Product status	Live

TL;DR

• Frontier is the quality ceiling — use it where a wrong answer is expensive.
• Keep it off the default path; route to it on validation failure or complexity.
• Use it to build the evaluation baseline other models are measured against.
• Send complete, well-structured context so the call is not wasted.
• Score escalations regularly so the tier stays where it earns its keep.

How it works, step by step

1. Define what "hard" means for your product: failure cost, ambiguity, input length.
2. Build an escalation eval set from your hardest real requests.
3. Run plugsky-frontier against your current production model on that set.
4. Add a validator so escalation fires on evidence, not guesswork.
5. Cap escalation attempts per request and log the escalation rate.
6. Keep a faster model as the default path and cache repeated premium outputs.
7. Review escalation triggers monthly and after model-card changes.

Try it yourself

Open the LLM cost calculator →

What the frontier tier is for

plugsky-frontier is the top capability tier in the catalogue, intended for requests where answer quality matters more than response time. Typical jobs are complex analysis, high-stakes drafting, difficult reasoning chains and synthesis across several documents. It speaks the same OpenAI-compatible chat surface as every other model, including streaming, function calling and JSON mode.

The frontier model guide covers the model itself; this article is about deciding when a request deserves it. Read the live model card at /models before hard-coding window or capability assumptions.

Escalation triggers that work

• Validation failure: a schema, test, citation or rubric check fails on the cheaper tier.
• Complexity flags: long inputs, multiple documents or reasoning-heavy task types.
• Premium workflows: contract review, financial analysis or anything customer-facing that must be right.
• Offline jobs: batch analysis where throughput does not matter but quality does.

Triggers should be observable and testable. "The first answer looked weak" is not a trigger; "the verifier returned false" is. Keep the escalation evidence compact and send the original request unchanged so context is preserved through the chain.

Cost and quality discipline

Frontier calls are the ones worth measuring most carefully. Build a baseline: what does your current model score on your hardest set, and what does frontier add? If the delta is small, the escalation rule — not the model — is the problem.

Cache outputs for repeated premium tasks, store them with the source inputs, and review the escalation rate monthly. Self-serve plans are flat monthly with fair-use usage rather than per-token billing, so use the live pricing page to map escalation volume to plan limits, and protect quality by keeping the frontier path well fed rather than overused.

Honest comparison

Dimension	Frontier tier	Workhorse tier	Small tier
Quality ceiling	Highest in the catalogue	Strong general purpose	Good for narrow, simple tasks
Latency profile	Slowest	Balanced	Fastest
Best use	Escalation and baseline	Default production traffic	Cheap first hop
Cost profile	Highest per call	Moderate	Lowest
Failure impact	Lowest risk	Manageable with validators	Needs escalation path
Failover	Same-profile top-tier peers	Peer fallback	Peer fallback