Skip to content

About

Helps Claude pick the right model and effort level for the work, based on benchmarks, current pricing, and your plan. Picks reviewed weekly.

Topics

Resources

Stars

1 star

Watchers

1 watching

Forks

Latest commit

 

History

15 Commits

Folders and files

Repository files navigation

model-picker

Use the right Claude model for the job. Ask Claude which model and effort level a piece of work needs, and get one clear pick with the reason behind it.

Current as of 2026-11-12. Covers Claude Fable 5.2, Opus 5.5, Sonnet 5.5 and Haiku 5.5.

The best model choices are reevaluated every week, so update the plugin now and then to get the latest picks.

What it does

Claude comes in several models, and each can run at several effort levels. The biggest model at the highest effort is rarely the right choice. It costs more, it is slower, and on Anthropic's own charts the top effort level often scores no better than the one below it. The smallest model is not the right choice for hard work either, because failed attempts cost more than doing it properly once.

This plugin gives Claude a tested way to choose. When you ask, Claude will:

  • Recommend a model and effort level for a task, based on what is at stake, what kind of work it is, and how you pay for Claude.
  • Explain the pick in a sentence or two, and tell you what to step up to if it falls short.
  • Set up agents in Claude Code so each kind of work runs on a suitable model, instead of everything running on your most expensive one.
  • Check what your agents really ran on, from the session records, so you are not guessing.
  • Assign a model and effort to every agent in a multi-agent build, in the plan or as agents are started.

Who it is for

Anyone who uses Claude for real work and wants good results without wasting money or plan allowance. It is most useful if you:

  • Use Claude Code with helper agents.
  • Are on a plan with weekly limits and want them to last.
  • Pay per token and want a lower bill.
  • Are not sure when Fable is worth it over Opus, or Opus over Sonnet.

Install

From the Claude desktop app

  1. In the left sidebar under Personal plugins, click the + button.
  2. Choose Create plugin, then Add marketplace.
  3. In the URL field, enter Pupfish-LLC/claude-plugins, then click Sync.
  4. Click the + button again and choose Browse plugins.
  5. Find model-picker and install it.

From the command line

  1. Add the Pupfish plugin marketplace, if you have not already:

    /plugin marketplace add Pupfish-LLC/claude-plugins
    
  2. Install the plugin:

    /plugin install model-picker@pupfish
    

How to use it

In a conversation

Just ask in plain words. There is nothing to memorize.

  • "Which model should I use to refactor this billing module?"
  • "Is Fable worth it for this, or is Opus enough?"
  • "I'm about to build this feature with several agents. What should each one run on?"
  • "My weekly limit keeps running out by Thursday. What should I change?"
  • "Set up my Claude Code agents so searches stop running on the expensive model."
  • "Check which model and effort my agents ran on in this session."
  • "A new Claude model just came out. Does it change what I should use?"

You can also call it directly with /model-picker.

What an answer looks like:

Pick: Sonnet 5.5, medium
Why: The change is fully specified and your tests will catch a miss, so
     this is scoped work that does not need a larger model.
If it falls short: raise effort to high before moving to Opus 5.5.
How to set it: /model claude-sonnet-5-5 and /effort medium at the start
     of the session.

For a project with several kinds of work, you get a small table with one row per role.

In a multi-agent build

A session that coordinates other agents can use the plugin in two ways. Add one of these to the project's instructions.

Simplest. Decide as agents are started.

Before starting any helper agent, use the model-picker skill to choose its model and effort.

For a planned build. Decide once, in the plan.

While planning, load the model-picker skill and give every work package an agent role, such as builder-critical or reviewer. Start agents by role, and pass no model or effort on the call.

The second needs the agent files set up first. It makes each choice once, in writing, where you can review it.

The first thing it needs to know

The right pick depends on how you pay, so Claude may ask one question:

  1. Fable is included in your plan's weekly limit. This is Max, Ultra, and premium seats on Team and on legacy seat-based Enterprise plans. Fable can use up to 50 percent of your weekly limit. It draws on the same weekly limit as the other models and uses it faster, so it is not a separate allowance. Opus 5.5 at high effort is the default for substantial work. Fable 5.2 is kept for the highest-stakes work, and for building work where Opus 5.5 clearly tried and was still wrong.
  2. You pay per token. This is a key from the Claude Console, a cloud provider, or a usage-based Enterprise plan. Fable 5.2 at medium costs about the same per task as Opus 5.5 at high, so Fable 5.2 at medium is the default for substantial work where Fable is offered.
  3. Fable bills to usage credits, or is not available. This is Pro, and standard seats on Team and on legacy seat-based Enterprise plans. It also applies when your Fable share is used up, under zero data retention, and where an Enterprise admin has not turned Fable on. Opus 5.5 at high is the default, with Fable saved for the hardest work. If you choose to pay for Fable with usage credits, you pay per token for that work, so case 2 applies to it.

Launch week only: until 2026-11-16 at 11:59 pm Pacific time, Pro plans and standard Team seats have Fable included for up to 50 percent of their weekly limit. They stay in case 3 that week. After it, expect Fable to bill to usage credits.

If you are not sure, open the /model picker in Claude Code. "Requires usage credits" on the Fable row means the third case. So does a Fable row that is missing or grayed out. No such label, together with a Fable bar on your usage page, means the first. On Pro and standard Team seats this check may not work during the launch week, so go by your plan name.

The Free plan is outside these picks. It has no Opus, no Fable and no Claude Code.

The plan rules are from Anthropic's help pages and the /model label is from the Claude Code documentation. Both were read on 2026-11-09 and again on 2026-11-12. The Free plan facts are from Anthropic's plan pricing page, read on 2026-11-09. The sources are in plans-and-costs.md.

The current picks at a glance

Three kinds of work depend on which case you are in.

The work Case 1: Fable included in your plan Case 2: you pay per token Case 3: Fable on usage credits, or not available
Highest stakes, where a wrong answer would be silent and costly Fable 5.2, xhigh Fable 5.2, xhigh Opus 5.5, xhigh
Substantial building and coding with tests Opus 5.5, high. If it skips checks or stops early, xhigh. If it tries and is still wrong, Fable 5.2, high Fable 5.2, medium. Step up to high Opus 5.5, high
Independent review of another model's work Opus 5.5, xhigh for highest-stakes work built on Fable, high for other work built on Fable. A fresh Opus 5.5 agent, high, for work built on Opus Opus 5.5, high A fresh Opus 5.5 agent, high

The rest is the same in every case.

The work Every case
Coordinating other agents Opus 5.5, high
Scoped, fully specified changes. Sifting logs Sonnet 5.5, medium
Research with sources Sonnet 5.5, high
Everyday questions and drafting Sonnet 5.5, default
Summarize, classify, extract at volume Haiku 5.5, medium
Read-only search and lookup Haiku 5.5, low

A few rules do most of the work:

  • Start from the stakes, not from the size of the task.
  • Tests come first. A model's review adds to them.
  • Have work reviewed in a fresh session, by a different model from the one that wrote it when you can.
  • Do not use the max effort level unless you have tested that it helps on your task.
  • Pick the model at the start of a long session. Switching partway is expensive.
  • Do not start a helper agent for something a single lookup can do.

What it is based on, and its limits

The picks come from Anthropic's published release tables, its charts of benchmark score against cost per task and its guide to cost and intelligence, plus the Claude Code documentation and changelog. Each number in the plugin is labeled as published by Anthropic, read from a chart, reported by a third party, or observed once in real use.

Please keep these limits in mind:

  • The figures are Anthropic's own, not measurements on your work. They are a well-supported starting point. If a choice matters to you, measure it on your own tasks.
  • Chart readings are approximate. Gaps of two or three points between two picks are within normal noise.
  • Two picks go further than Anthropic's advice. Anthropic suggests starting most work on Opus 5.5 at its default effort, medium, and moving up to Fable 5.2 for demanding, long-running work or when Opus 5.5 at higher effort still falls short.
    • Fable 5.2 at medium when you pay per token. The plugin makes it the default for substantial building, because on two coding benchmarks it scores about level with or slightly above Opus 5.5 at high for about the same cost per task. One of those is a chart reading and the other is a single published run. The choice is close. Two other benchmarks point different ways, so Opus 5.5 at high is a sound alternative.
    • High effort for building on Opus 5.5. The plugin uses high where Anthropic starts at medium. On the five benchmarks that show both, high scored about 2 to 7 points above medium for about 20 to 65 percent more cost per task. The gain is small. If your tests reliably catch failures, and waiting for a second run on those failures is acceptable, Anthropic's own figures on one benchmark favor starting at medium and rerunning only the failures at high.
    • On a plan that includes Fable, the plugin defaults to Opus 5.5 for building. There Fable draws on your one weekly limit faster than the other models do, and Anthropic has not published how much faster.
    • These are the plugin owner's choices. To follow Anthropic's advice as given, start on Opus 5.5 at medium.
  • Plan rules change often and some are not published. The plugin says which ones it could not confirm.
  • It only covers Claude models.

The full evidence, with sources and dates, is in skills/model-picker/references/evidence.md.

How it stays current

Model lineups, prices and published results change every few weeks. The best model choices are reevaluated every week against Anthropic's latest releases, prices and documentation. A new version is published when:

  • Anthropic releases a new model or retires one.
  • Anthropic changes prices, plan limits or how Claude Code chooses models.
  • New published evidence, or measurements on real work, change which pick is best.

Every change to the guidance is reviewed by a person before it is published. In a week with no change, a new version is still published with that week's review date, so the date at the top of this page tells you when the picks were last checked. Every version states the date it was last reviewed. If that date is more than about two months old, or you ask about a model the plugin does not know, Claude will tell you the guidance may be behind.

Getting the latest version

You only receive a new version when you update.

  • Automatic. Run /plugin, open Marketplaces, select pupfish, and choose Enable auto-update.
  • By hand. Run /plugin marketplace update pupfish in a session. In the desktop app, sync the marketplace again the same way you added it.

Changes are listed in CHANGELOG.md. The update process is described in MAINTAINING.md.

Setting up agents

Optional. If you let Claude hand work to helper agents, the plugin comes with ready-made agent files, so that searches, reviews and small jobs stop running on your session's most expensive model. Ask Claude to "set up model-picker agents" and it walks you through it. Nothing is copied into your projects until you ask. The details are in Setting up agents.

Feedback

If a pick did not work for you, or your own measurements disagree with the guidance, please open an issue at https://github.com/Pupfish-LLC/model-picker/issues. Real measurements are the most useful thing you can send.

License

Released under the Apache License, Version 2.0. You are free to use it, change it and share it, including in commercial work. If you pass it on or build on it, keep the NOTICE file with it and say which files you changed. The license does not grant use of the Pupfish name.

Author

Built by Pupfish Analytics. Part of the Pupfish Claude Marketplace.

This plugin is not affiliated with or endorsed by Anthropic. Claude and the model names are trademarks of Anthropic.

About

Helps Claude pick the right model and effort level for the work, based on benchmarks, current pricing, and your plan. Picks reviewed weekly.

Topics

Resources

Stars

1 star

Watchers

1 watching

Forks

Contributors