Key facts
| High-volume tier | plugsky-lite and plugsky-micro — free plan, no card required |
| Quality tier | plugsky-plus and plugsky-pro for tone-sensitive and complex replies |
| Escalation tier | plugsky-max for policy analysis and hard cases |
| Tool calling | OpenAI-style function calling is live on supported models |
| Capabilities | Streaming, JSON mode and function calling on the chat completions endpoint |
| Free plan | Two free models, plugsky-micro and plugsky-lite |
| Trial | 14-day full-access trial for evaluating paid tiers |
| Product status | Live |
TL;DR
- Triage and FAQ answers: start on the free models, plugsky-micro and plugsky-lite.
- Tone-sensitive replies: move to plugsky-plus or plugsky-pro.
- Reserve plugsky-max for policy analysis and escalation summaries.
- Give the model tools for order and account lookup — never invent customer data.
- Score candidates on your own transcripts and keep the cheapest passing tier as default.
How it works, step by step
- Classify your support intents and decide which are safe to automate fully.
- Run FAQ, tagging and drafting prompts on the free tiers first.
- Add plugsky-plus or plugsky-pro for complaint handling and tone-sensitive flows.
- Connect tools such as order lookup and account status, and validate arguments server-side.
- Add guardrails: allowed tools, refusal rules, PII redaction and a human handover path.
- Score every tier on your own transcripts and keep the cheapest passing tier as the default.
- Log model, prompt and tool call per conversation for audit and continuous tuning.
Try it yourself
Match the tier to the support moment
Support work is a sequence of moments, and each one has a different quality bar. Triage — intent detection, routing, tagging — is high volume and low risk, which makes it a natural fit for the free tiers. Resolution is where tone and judgement matter: refunds, complaints and account issues usually deserve plugsky-plus or plugsky-pro. Escalation, policy questions and multi-conversation summaries are the rare, high-cost moments where a top tier such as plugsky-max earns its place.
Because every tier shares the same OpenAI-compatible endpoint, you can start on plugsky-lite and plugsky-micro, then change the model name per flow as quality requirements grow — no rewrite required.
Tools, guardrails and handover
A support model without tools can only talk; with tools it can act. Function calling lets the model request an order lookup, a subscription status or a returns check, and your code executes it. Keep the safety boundary in your application:
- Allowlist tools per flow and validate every argument before execution.
- Never let the model fabricate account data — it must retrieve or say it cannot.
- Redact PII from prompts where it is not needed, and log what was sent.
- Require human approval for refunds, credits or anything irreversible.
- Define an explicit handover phrase and route to a person when confidence is low.
The function calling API guide covers the request shape and loop in detail.
Evaluating support quality
Ranking models on generic benchmarks tells you very little about support quality. Build a transcript set from real conversations, label the ideal outcome, and score candidates on correctness, tone, refusal behaviour and whether they avoided inventing information. Include the ugly cases: angry customers, ambiguous requests, mixed languages and out-of-policy asks.
Then compare tiers on the same set and route by intent: the cheapest tier that passes a flow becomes the default, with a stronger model behind it for failures. Re-run the evaluation after model-card changes, and treat a regression as a routing bug. Pricing is flat monthly on self-serve plans — see the live pricing page for current plans rather than estimating per-token costs.
Honest comparison
| Support job | Free tier (micro, lite) | Paid mid-tier (plus, pro) | Top tier (max) |
|---|---|---|---|
| FAQ and macro answers | Best fit | Strong | Overkill |
| Intent classification and tagging | Best fit | Strong | Overkill |
| Tone-sensitive complaints | Drafts only | Best fit | Good for hard cases |
| Order or account lookup | Tool calls supported | Best fit | Works |
| Escalation and policy analysis | Not recommended | Good | Best fit |
| Trial before committing | Free plan, no card | 14-day full-access trial | 14-day full-access trial |
Frequently asked questions
Which Plugsky model should a support bot use by default?
Use plugsky-lite or plugsky-micro for triage, FAQs and tagging, and route complaints and complex replies to plugsky-plus or plugsky-pro. Keep plugsky-max for escalation summaries and policy questions.
Can the model look up orders and accounts?
Yes — with function calling, the model requests a tool and your code returns the data. Validate arguments server-side and never let the model invent customer information.
Does the free plan work for support automation?
The free plan includes plugsky-micro and plugsky-lite with no card required, which is enough to build and evaluate triage and FAQ flows before moving paid flows into production.
How do I stop the bot inventing policies?
Ground every answer in retrieved documents or tool results, add explicit refusal rules, and require a citation or tool call before any policy statement. Test refusal behaviour on adversarial transcripts.
How should I evaluate support models?
Score candidates on real transcripts for correctness, tone, refusal behaviour and hallucination, then route each intent to the cheapest tier that passes. Re-run the evaluation when model cards change.
Is there a trial for the paid tiers?
Yes — the 14-day full-access trial lets you evaluate paid tiers on your own support flows. See the live pricing page for current details.
How is pricing structured for support workloads?
Self-serve plans are flat monthly with fair-use usage rather than per-token billing, which keeps high-volume support traffic predictable. Check the live pricing page for current plans.