Key facts
| Step 1 | Benchmark plugsky-micro and plugsky-lite on your real prompts |
| Step 2 | Choose plugsky-plus or plugsky-pro when paid tiers are needed |
| Step 3 | Reserve plugsky-max for hard analysis and escalations |
| Specialists | Long-context, vision and coding tiers cover specific workloads |
| Catalogue | 30+ models behind one OpenAI-compatible endpoint |
| Policy | Keep one default and one escalation model per workload |
| Free plan | plugsky-micro and plugsky-lite, no card required |
| Product status | Live |
TL;DR
- Start free and let evaluations — not vibes — decide when to move up.
- Reasoning-heavy tasks: plugsky-micro first; speed: plugsky-lite.
- Paid ladder: plugsky-plus, then plugsky-pro, then plugsky-max for hard cases.
- Reach for specialists for long context, vision and coding.
- One default plus one escalation model per workload keeps policy manageable.
How it works, step by step
- Write down the workload, its output shape and what a failure costs.
- Run plugsky-micro and plugsky-lite against a labelled evaluation set.
- If they pass, ship on them; if not, note exactly where they fail.
- Pick the cheapest paid tier that fixes those failures — plugsky-plus or plugsky-pro.
- Add a specialist only for a real capability gap such as long context, vision or coding.
- Set plugsky-max as the escalation target for the hardest cases.
- Keep the decision in configuration and re-run the evaluation when model cards change.
Try it yourself
Three questions before you pick
Model selection is a short interrogation. What is the task? Chat, extraction, summarisation, code, vision and retrieval all point at different tiers. What does failure cost? A mislabelled ticket is cheap; a wrong contract summary is not. How much volume? High-volume simple work rewards small fast models, while rare complex work can afford a stronger tier.
Answer those three and half the catalogue eliminates itself. The rest is decided by evaluation, not opinion.
Mapping workloads to tiers
Everyday chat and FAQ: start with plugsky-micro or plugsky-lite, move to plugsky-plus if you need vision or stronger answers. Extraction and tagging: the free tiers with a JSON schema validator. Long documents: a long-context tier such as plugsky-longctx or plugsky-kimi. Images: plugsky-vision-fast, plugsky-qwen-vl or plugsky-gemma-4. Coding: plugsky-coder-fast for small diffs and plugsky-coder for multi-file work. Hard analysis: escalate to plugsky-max.
Check the live model card at /models for the current window, capability flags and routing before locking anything in.
Make it a policy, not a favourite
Teams drift toward one favourite model because it worked once. A policy is more durable: one default and one escalation model per workload, stored in versioned configuration, with an evaluation set that runs when model cards change.
That structure also makes cost conversations concrete — you can see which workloads escalated and why, rather than debating model preferences. Self-serve plans are flat monthly with fair-use usage rather than per-token billing, so use the live pricing page for plan limits and keep the evaluation focused on quality and routing. For a wider view of the architecture, see the OpenAI alternative guide.
Honest comparison
| Workload | Free start | Paid step-up | Specialist option |
|---|---|---|---|
| Everyday chat | micro or lite | plus | Not needed |
| Extraction and tagging | lite with JSON mode | pro | Not needed |
| Long documents | lite for short inputs | pro | longctx or kimi |
| Images and screenshots | Not covered | plus (vision) | vision-fast or qwen-vl |
| Coding | micro or lite for snippets | pro | coder or coder-fast |
| Hard analysis | Not recommended | pro | max |
Frequently asked questions
Which Plugsky model should I start with?
Start with plugsky-micro if the task benefits from reasoning and plugsky-lite if speed and volume matter. Both are free with no card required, which makes them the natural baseline.
When should I move to a paid model?
When your evaluation set shows the free tiers failing in a way that matters — depth of reasoning, vision inputs or long context. Then pick the cheapest paid tier that fixes those specific failures.
How do I choose between plugsky-plus and plugsky-pro?
plus suits everyday work with vision and tools; pro suits broad production workloads that mix reasoning, code and agents. Evaluate both on the workload before deciding.
When do I need a specialist model?
When the requirement is a specific capability rather than general quality: a long-context tier for many-page inputs, a vision tier for images, or a coding tier for diffs and refactors.
How do I know the choice is still correct?
Keep an evaluation set per workload and re-run it when model cards change. A decision that was right last quarter can drift as models, prompts or traffic change.
Can I change models later?
Yes — every model shares the same OpenAI-compatible endpoint, so switching is a model-name change. Keep the name in configuration rather than hard-coded across the codebase.
How is pricing structured?
Self-serve plans are flat monthly with fair-use usage rather than per-token billing. Check the live pricing page for current plans and limits.