Models + Cost

Which Plugsky model is best for coding?

Pick the coding model in Plugsky's 30+ model catalogue by task, not by a single leaderboard. plugsky-coder-fast suits tight edit-and-test loops and small diffs, plugsky-coder handles multi-file refactors and reviews, and plugsky-coder-ultra is the specialised code tier for focused generation. A general workhorse such as plugsky-pro covers sessions that mix code, tool calls and prose. Benchmark candidates on your own repository and test suite before standardising.

Key facts

Fast coding tierplugsky-coder-fast for autocomplete, small diffs and edit loops
Deep coding tierplugsky-coder for multi-file reasoning, refactors and review
Specialised tierplugsky-coder-ultra for focused code generation and explanation
General workhorseplugsky-pro for sessions that mix code, tools and prose
CapabilitiesStreaming, function calling, JSON mode and long-context on the OpenAI-compatible API
Context class128K-class windows on current coding models; live limits are published per model
Free planplugsky-micro and plugsky-lite, no card — enough for small edits and tests
Product statusLive

TL;DR

  • Match the tier to the task: fast for small diffs, deep for multi-file work.
  • plugsky-coder-fast keeps agent edit-test loops cheap.
  • plugsky-coder and plugsky-pro handle refactors and mixed code plus prose.
  • plugsky-coder-ultra is the focused code-generation specialist.
  • Your repository test suite is the only coding benchmark that matters.

How it works, step by step

  1. List the coding tasks you actually ship: autocomplete, agent edits, refactors, reviews and tests.
  2. Score plugsky-coder-fast, plugsky-coder and a general workhorse on a fixed set of real tasks.
  3. Use your repository test suite as the acceptance gate instead of sample snippets.
  4. Measure cost per accepted change, not per request, and note which tier wins each task.
  5. Encode the result as routing rules: small diffs to the fast tier, repeated failures to the deep tier.
  6. Add a fallback so a failed edit retries once on a stronger model before a human reviews.
  7. Re-run the evaluation when the live model card changes.
1List the codingtasks you actuallyship: autocomplete,2Scoreplugsky-coder-fast,plugsky-coder and a3Use your repositorytest suite as theacceptance gate4Measure cost peraccepted change,not per request,5Encode the resultas routing rules:small diffs to the6Add a fallback so afailed edit retriesonce on a stronger

Try it yourself

Open the coding model selector →

The coding tiers and what each is for

plugsky-coder-fast is built for high-frequency, low-diff work: inline completions, small patches and agent edit-test loops where latency and cost per attempt matter. plugsky-coder is the deeper tier for multi-file reasoning, refactors and review comments that require holding more context. plugsky-coder-ultra is the specialist for focused generation of self-contained functions and modules.

Keep one general workhorse (plugsky-pro) for sessions that mix code with tool calls, shell commands and prose explanations. Switching between tiers is a model-name change on the same endpoint, so mixing them inside one product is cheap to implement.

Choosing for agentic coding workflows

In an agent loop, the model proposes an edit, your harness applies it, runs tests and feeds failures back. That loop changes the selection criteria: cost per accepted change and recovery behaviour matter more than one-shot accuracy.

  • Start the loop on the fast tier; escalate to a deeper tier after one or two failed attempts.
  • Cap iterations and diff size so a weak attempt cannot burn the whole budget.
  • Send the exact failing test output back as context — it is the highest-signal input you have.
  • Use function calling for file and test operations instead of parsing free-form text.
  • Keep a deterministic linter or formatter in the loop so trivial issues never consume model calls.

Evaluating coding models on your repository

Public coding benchmarks do not know your framework, conventions or test suite. Build a private evaluation set from merged pull requests: prompt the model with the pre-change file plus the issue text, apply the candidate patch, and score it with your real tests. Record first-pass rate, retries needed and the model that ultimately succeeded.

Then map the results to routing rules rather than a single winner. Most teams end with a fast default for small changes, a deep tier for large diffs and review, and an escalation path that fires only when tests fail. See model routing for how to implement the policy without hard-coding model names throughout the codebase.

Honest comparison

Coding taskFast tier (coder-fast)Workhorse tier (coder, pro)Specialised tier (coder-ultra)
Autocomplete and small diffsBest fitWorks, higher costOverkill
Multi-file refactorRisky on large diffsBest fitGood for focused modules
Test generationGood for simple casesStrong all-roundGood for isolated units
Review and explanationFine for style checksStrong with long contextGood for focused review
Agent edit-test loopsBest cost per loopUse after failed attemptsSituational

Frequently asked questions

Which Plugsky model is best for coding?

Use plugsky-coder-fast for small diffs and tight loops, plugsky-coder for multi-file work and reviews, and plugsky-coder-ultra for focused generation. Evaluate them on your own repository before standardising.

Can I use the free plan for coding?

Yes. The free plan includes plugsky-micro and plugsky-lite with no card required, which is enough for small edits, test scaffolding and experimentation before you move to paid coding tiers.

Does the API support function calling for code tools?

Yes — OpenAI-style function calling is live on supported models, so an agent can call file, shell and test tools through the standard chat completions shape.

What context window do the coding models have?

Current coding models are 128K-class, but the live window and output limits are published per model on the model catalogue. Use that page as the source of truth.

How should I benchmark coding models?

Build a private set from real pull requests, apply each model's patch, and score it with your own test suite. Track first-pass rate, retries and cost per accepted change rather than raw token throughput.

Can I route between coding models at runtime?

Yes. All tiers share the same OpenAI-compatible endpoint, so routing is a model-name decision inside your agent harness — see the model routing guide for patterns.

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.

Is there a trial for the paid coding tiers?

Yes — a 14-day full-access trial lets you evaluate paid coding tiers on your own tasks before committing. See the live pricing page for details.