rightsize

A Claude Code mod that matches each prompt to the right model and effort level, so you don't pay Opus prices for Haiku work.

Created by Archi Dubey. MIT license.

One demo run in the Express repository. rightsize's choice depends on the task, your model and your settings, so yours may differ.

Before you use it

rightsize is a work in progress. It changes how Claude Code spends your tokens, and its choices are estimates. Savings are not guaranteed: some turns can cost more, not less, and a cheaper model can sometimes do weaker work. Use it at your own risk, and check /rightsize report to see what it really did for you.

Give it a few turns to learn. rightsize starts with general estimates. It needs some turns of your real work before it knows how your prompts, your codebase, and your models behave. Don't expect the first few recommendations to save money, and some may cost a little more. As it compares each estimate with what you actually spent, its calls get better. Judge it after a few sessions, not a few prompts.

Feedback and feature requests are welcome. If you find a bug, a wrong call, or a saving that didn't show up, or you have an idea for a feature, open an issue. For anything you'd rather not post in public, email [email protected]. What you report shapes what gets built next.

What it does

Not every prompt needs your strongest model thinking its hardest. "Rename this variable" and "redesign the auth flow" get the same treatment by default. rightsize looks at each prompt before it is sent and picks a better fit.

It sizes each prompt

• Before each prompt goes out, a quick, cheap check rates how hard the work is.
• Simple work runs at lower effort. Hard work gets more.
• If Claude gets stuck partway through (the same command keeps failing, or it goes in circles), rightsize turns the effort up for the rest of that turn.

It hands work to cheaper helpers

• When a smaller model can do a task well, rightsize sends the task to a helper (a subagent) on that model. Your main chat stays as it is.
• The main chat writes the helper a clear brief, checks the result, and asks for fixes if needed. If the helper still can't get it right, the main chat does the work itself.
• When Claude starts its own helpers, rightsize moves them to a cheaper model too, if that model can do the job.

It looks after your budget

• On a subscription, it watches your 5-hour and weekly limits. If you are using them up too fast, it saves harder. If you will have plenty left over, it lets Claude work harder.
• It warns you before an old, idle session would cost you a lot to pick up again, and shows cheaper options. On a subscription it gives the cost as a share of your 5-hour window.
• It keeps its price list up to date with Anthropic's pricing page.

It learns and shows its work

• After each prompt, it compares what it expected to spend with what you really spent, and it adjusts its guesses next time.
• /rightsize report opens with one sentence on how much it saved in the last 30 days, then a table of what it changed and saved this session, in the last 7 days and in the last 30 days. Below that it lists which turns it did not count as savings and why.
• The report also shows the fixed context that comes with every message (tools, instructions, plugins), largest part first, and the next steps to cut it.

You stay in charge

• The first time each kind of change comes up, rightsize asks you what to do: do it automatically, ask each time, or never. You can also set your answer ahead of time in /config.
• A small bar above the prompt tells you what it changed, with buttons to undo it or change your rule.
• Every limit and feature can be changed or turned off in /config.
• Your setting wins. A value you set yourself is never changed or bypassed by rightsize's automatic behavior. Its automatic adjustments only act on settings you left at their default.

Install

You need Claude Code 2.1.287 or newer in the terminal, or 2.1.286 or newer in the desktop app. Installing takes about 2 minutes.

Install it yourself

1. Open a terminal (on a Mac, the Terminal app).

2. Check your Claude Code version:

   claude --version
   

   If the number is lower than 2.1.287, update Claude Code first:

   claude update
   

3. Add the rightsize marketplace, so Claude Code knows where to find it:

   claude plugin marketplace add archid21/rightsize
   

   It should end with Successfully added marketplace: archid21.

4. Install rightsize for all your projects:

   claude plugin install rightsize@archid21 --scope user
   

   It should end with Successfully installed plugin: rightsize@archid21 (scope: user). A note that some options are not set yet is normal: every option has a default.

5. Start a new Claude Code session (type claude, or open a new session in the desktop app). Sessions that were already open don't load it.

6. Type /rightsize and press Enter. If it shows its status, it works.

Already inside Claude Code? You can type /plugin marketplace add archid21/rightsize and then /plugin install rightsize@archid21 instead of steps 3 and 4. When it asks where to install, pick the user scope.

Ask Claude to install it

Paste this into any Claude Code session:

Install the rightsize plugin from GitHub for me. Run these two commands in the terminal:
1. claude plugin marketplace add archid21/rightsize
2. claude plugin install rightsize@archid21 --scope user
Tell me if either one fails and why. When both succeed, tell me to start a new session and type /rightsize to check that it works.

Claude asks your permission before it runs each command. When it is done, start a new session and type /rightsize.

The four questions

Each question is asked once, the first time its case comes up:

Case	Choices
Q1 A prompt needs less effort	Lower automatically · Ask me each time · Never lower
Q2 It needs more effort on the same model	Raise automatically · Ask me each time · Never raise
Q3 A stronger model would do better	Ask before sending it to a subagent · Send it automatically · Stay on this model, raise effort only · Ask me each time
Q4 A task could run on a cheaper model	Send it to a cheaper subagent · Ask before sending it · Stay on this model, lower effort only · Ask me each time

After your first answer, rightsize asks whether to use your choices in every session. Each question is also a row in /config (Lowering, Raising, Stronger model, Cheaper model), and Remember choices answers the follow-up ahead of time. A rule you set in a row is used as it is, and its question is never asked.

How it decides

• Your setting wins. A value you set in /config, or answered when rightsize asked, is never changed or bypassed by rightsize's own behavior. That covers the thresholds, the four rules, Reports, and every other row. Budget pacing moves only the thresholds you left at their default. A window nearly used up skips raises and stronger models only while their rule has no answer. Rate limits never change Reports once you set it; /rightsize says when they disagree. A value counts as yours once you set it, even if you set it to the number that is the default. /rightsize lists the settings you changed; the rest are defaults, shown in /config.

• Effort changes can cost a cache rewrite. Anthropic's docs say that changing effort at the top level of a request resets the conversation's prompt cache. A per-message change (beta) keeps it. rightsize counts that cost, and it measures what Claude Code really does: every effort switch is logged with whether the cache survived. Lowering effort happens only when it pays.

• Changing the model always means a subagent. The main chat keeps its model and its cache. A hidden note beside your prompt (your text is unchanged) asks it to:

  1. write a brief with a "Done means" list
  2. check the subagent's result against that list
  3. send at most 2 rework notes
  4. tell you how it went

  When the work goes to a cheaper model and still fails after 2 reworks, the main chat finishes it itself.

  Steps only you can approve stay in the main chat. A hand-off's subagent must not push, publish, release, deploy, send messages or act in your accounts, because a message from another agent is not your approval. Its brief tells it to prepare the work (edits, tests, notes) and stop, and rightsize blocks any push, publish, release, deploy or send command it tries anyway. The main chat checks the work and does those steps itself. A request that is mainly such a step ("push and release it") isn't handed off at all: the Haiku rating marks it, and only effort can change. When Claude keeps a chosen hand-off in the main chat because it pushed, published, deployed or sent something, the band says so and /rightsize report doesn't count it as a skipped hand-off.

• The subagent runs at the level its model needs. rightsize sets the effort of the hand-off's subagent itself, so the level does not depend on Claude passing it on.

• Claude's own subagents are right-sized too. When Claude starts a general-purpose subagent without naming a model, it would run on Claude's model (often Opus). rightsize rates the subagent's brief with Haiku and runs it on the cheapest weaker model that fits, at that model's level, when that costs under the hand-off threshold. It never moves a subagent to a stronger model. It leaves alone subagents that name a model, other agent types (they may set their own model, which rightsize can't see), forks, workflow agents, resumes, teammates, and subagents started by other subagents. Each one it moves costs one Haiku rating call. in turns this off.

Choosing a starting model

rightsize works from the model and effort level your session starts on. It can lower effort, hand work down to a cheaper model, or hand it up to a stronger one, so the start changes what you spend.

Start on Sonnet 5.5 at medium or Opus 5.5 at medium.

• Sonnet 5.5 at medium costs the least when most of your work is lookups, searches, edits and writing. Work that needs Opus goes to an Opus helper, and the Sonnet main chat checks it.
• Opus 5.5 at medium is the safer start when most of your work is features, debugging or design. rightsize then only hands work down, and Opus checks each result. Anthropic's benchmarks put Sonnet at high effort about 5 points behind Opus at medium on coding tasks.

Avoid starting on Opus at low, which makes rightsize raise effort often (each raise re-writes the whole conversation to the cache), or on Fable. A high start gives rightsize more to save, but the total is still the highest: a big saving percent means a high start, not a cheap one.

rightsize tells you when another start would suit you better. Once a week (or once a month: Start suggestions in /config), at the start of a session, it estimates what your last 30 days of rated prompts would have cost from Sonnet and Opus at medium and at high. Every rated prompt records the level each model would need, so the same prompts can be estimated from any start, with rightsize's hand-offs and effort changes counted. The band suggests a switch when:

• another start is at least 15% cheaper (25% while rightsize has measured few runs on it), and it would send no more than 1 prompt in 10 to a stronger model; or
• more than 15% of your prompts needed a stronger model, or Claude got stuck and effort went up mid-turn. Then it suggests a stronger start.

Where Haiku's rating of a model has been wrong for a task type (see How it decides), that share of those prompts is estimated as if the model had been rated unable, so a start that leans on a model Haiku keeps misjudging looks dearer and sends more work up. The estimate follows your Use Haiku, Use Fable and Highest effort settings, so it never counts work sent where you said no. It waits for 50 rated prompts in 30 days, and 20 on your current start. Not now quiets the same suggestion for 14 days; Stop suggesting turns it off. /rightsize report always shows the comparison. rightsize never changes your model itself: switch with /model, which also sets the effort level.

The band

A one-row band above the prompt shows what rightsize did, with buttons to act on it:

rightsize: Effort lowered: high → medium for this prompt (automatically).  [Undo] [👍] [👎] [Ask me each time] [Never lower] [⋯] [×]

• Undo puts the prompt back in the box, to run at your normal level. Clear drops a session-wide switch.
• The rule buttons change your answer for this case (Q1–Q4) in one press, and set the matching row in /config too. Your current answer is not shown.
• ⋯ chooses when the band shows. × closes it until your next prompt.
• In the CLI, focus the band (ctrl+x tab or a click) and press a button's number.
• 👍 and 👎 appear on a hand-off's result and on a lowering, so you can say whether it was the right call. Pressing one writes a feedback line to the run log, takes the buttons off the band so they can't be pressed twice, and shows a short thanks. Nothing is sent anywhere: the mark tunes rightsize's choices on this machine (see How it decides).
• After a hand-off, the band shows the result, for example Sent to Haiku: $0.04 vs about $0.31 staying on Fable (estimated), saved about 87%. With the band off, that result is a line in the transcript instead.

The band shows in the CLI and the desktop app. VS Code and the mobile app keep the status line.

Commands and settings

• /rightsize: the settings you changed (the rest are defaults, shown in /config), the session switch, the measured runs and the cache-switch results.

• /rightsize help: every command and what it does. An unknown command shows the same list.

• /rightsize report: what rightsize did for you. In the terminal it is plain text; the desktop app renders it as Markdown, so the tables and the tree go in code blocks there to keep their columns lined up. Each section is below, in the order it appears. A section with nothing to show is left out.

  • Headline: one sentence for the last 30 days. On a subscription it leads with the share of your 5-hour window that the saving kept, and about how many minutes that is at the current pace. On the API it leads with the dollars saved.
  • Table: columns for this session, the last 7 days and the last 30 days. Savings come first: Total saved (after the cost of the Haiku rating calls), the share of your 5-hour and 7-day windows saved (subscription), the share of cost saved, and Extra spent on harder prompts (what raising effort or using a stronger model cost, for better answers; it isn't taken off Total saved). Then what rightsize did: prompts rated, how many were kept, lowered or raised, how many were sent to a cheaper or stronger model, suggestions you declined, Claude's subagents that ran on a cheaper model, and close calls (how many the task-type guide decided, and how many took the other model to measure it; 7 and 30 days only).
  • Starting point: what your last 30 days of rated prompts would have cost from Sonnet and Opus at medium and high. A suggestion to switch goes under Next steps (see Choosing a starting model).
  • Fixed context: the tokens every request carried that rightsize can't trim, with a note that it adds to your cost in every session and cutting it saves more, what re-reading them cost at least and its share of what those turns cost, then a tree with the largest part first (MCP tools with the biggest servers named, skills and commands, agent descriptions, CLAUDE.md and memory).
  • Accuracy: how often Haiku's judgement was wrong, counted from rework, lowered turns that stepped back up, mid-turn raises, 👎, and Sonnet checks that disagreed; how many Sonnet spot-checks agreed with Haiku; the hand-downs paused for a task type once its rate reaches the Pause after misses line; and how far the estimates were off, early against late.
  • Not counted as savings: a count and a plain reason for each kind of turn that gets no saving or went off course. Lowered turns that stepped back up, turns raised mid-turn because Claude looked stuck, hand-offs that cost more than staying or needed rework, hand-offs Claude skipped (not counting those kept in the main chat for steps only you can approve), effort switches that missed the cache, turns you stopped, and turns that ran during or after a workflow.

Turn it off or remove it

Pause it. Type /mods and switch rightsize off. Claude Code then works as if rightsize weren't there, and nothing is lost. Switch it back on the same way. To turn off only one feature, such as hand-offs or the band, use its row in /config.

Disable it, keeping your settings and what it has learned:

claude plugin disable rightsize@archid21

New sessions start without it. claude plugin enable rightsize@archid21 turns it back on where it left off.

Uninstall it:

1. If you turned on Share anonymous feedback, type /rightsize and copy your install ID first. Uninstalling stops all sending, but data already shared stays on the server until you ask for it to be deleted: email [email protected] with that ID.

2. Remove the plugin. This also clears the settings you saved in /config:

   claude plugin uninstall rightsize@archid21
   

3. Optional: remove the marketplace entry too:

   claude plugin marketplace remove archid21
   

4. Optional: delete what rightsize left on your disk. Uninstalling keeps its stored data (your answers, calibration and counters) and its logs:

   rm ~/.claude/plugins/store/rightsize_*.json ~/.claude/mods-logs/rightsize*
   

   Skip this step if you might come back: reinstalling picks up the stored data where it left off.

Logs

These are in ~/.claude/mods-logs/:

• rightsize.log: each decision.
• rightsize-runs.jsonl: each rated turn, changed or not and stopped or not. Each line has the session ID, the change made and the change suggested, the estimated and actual cost (and whether the actual cost came from the ledger or from tokens), the rating call's cost, the rate-limit windows before and after, the turn's duration and its subagent calls, each model's rated level (fit), the rater's guesses of the tokens read and written and the tool calls (read, write, steps), for a hand-off the level its subagent ran at (subLevel), Claude's own subagents it moved to a cheaper model (resized), how a close call was decided (tie, also on each resized entry), the usage pace when it moved the thresholds (pace), mid-turn raises with their step and reason (escalations), main-chat requests (requests), the fixed context every request carried and what re-reading it cost (fixed), and its workflow calls (workflows, and afterWorkflow once one has run in the session). A rating that came to nothing is logged as an overhead line. Each line has an id. Every 50th rating also gets a check line: Haiku's fit, Sonnet's fit (under check) and Sonnet's cost. A 👍 or 👎 gets a feedback line with the run's id as ref, and value up or down. Each of Claude's subagents it moved to a cheaper model gets a subagent line when it answers: its models, level, real cost and the raw estimates on its model and on the parent's. /rightsize report reads this file.
• rightsize-switches.jsonl: each measured effort switch, and whether the cache was kept.

Privacy

rightsize keeps its logs on your machine. It sends nothing to its author unless you say yes.

What the rating sees. To rate each prompt, rightsize asks Claude Haiku through Claude Code. The rating call carries your prompt (up to its first 6,000 characters) and the end of Claude's last reply (up to 1,500 characters), and goes to Anthropic, just as your prompts do when you use Claude Code. When Claude starts its own helper, rightsize rates the helper's brief the same way. Every 50th rating is also sent to Sonnet as a check (Rater check in /config). None of this goes to rightsize's author.

Sharing is opt-in. After 20 rated prompts, rightsize asks once: "Help improve rightsize? Share anonymous numbers with its author." If you say no, or close the question, nothing is ever sent. You can change your answer any time with Share anonymous feedback in /config.

What is sent when you say yes, at most once an hour, only for what is logged after you said yes:

• For each rated prompt: the task type, the models and effort levels (rated, used and suggested), the change made, the rater's guesses of tokens read and written and tool calls, the conversation's size, the estimated and real cost, the rating call's cost, how many subagent calls and requests it took, whether it stepped back up, was raised mid-turn (a count, not the reason), was stopped or ran after a workflow, how a close call was decided, and how long it took. Each event carries the time it happened, and the server saves the time it received the batch.
• Each 👍 or 👎, linked to its prompt by a one-way code.
• Each Sonnet spot-check: the levels Haiku and Sonnet gave each model, and its cost.
• Your rightsize version, whether you pay by API or subscription, and a random install ID that is not made from anything on your machine.

What is never sent: your prompts, Claude's replies, code, file names, paths, tool and MCP server names, the reasons Claude got stuck, session IDs, or anything that names you. The server also rebuilds every event from a fixed list of fields and drops anything else.

Turning it off stops all sending at once. Turning it on again later starts a new install ID, so the old and new data can't be linked. rightsize also sends nothing while Claude Code's CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC is set.

The numbers go to feedback.aivalue.blog, a server the author runs, and are kept in a private store that only the author can read. They are used to tune rightsize's defaults, such as the hand-off threshold, the task-type guide and the estimates of work per model. Like any web server, it records the IP address each request comes from in its request logs, which the host (Vercel) keeps for a limited time; the IP address is not saved with the numbers.

Questions about shared data, or a request to delete it: email [email protected] with your install ID (/rightsize shows it while sharing is on).

Limits

• Workflows (ultracode). When Claude runs a workflow, it fans the work out to many agents. Workflows in /config says what rightsize does: by default (and until you answer the question it asks before the first workflow) it stops handing work off for the rest of that session; follow my rules lets hand-offs go on; right-size every agent also moves workflow agents to cheaper models. In every case later turns teach calibration nothing and claim no saving, since a workflow keeps running in the background and its cost lands in later turns. It still rates prompts and changes effort. It can't see that ultracode is on until the first workflow runs. /rightsize report lists these turns on their own line.

• The app's effort picker does not move. The status line shows the level in use.

• The rating is an estimate from Haiku. Calibration corrects cost guesses, but it can't correct a wrong choice of model. rightsize does measure how often the rating was wrong and stops trusting it where it keeps missing (see How it decides), but that takes enough outcomes: until they are in, a wrong call can still cost a rework or a step back up.

• Model facts (names, default effort levels, which models can switch effort mid-conversation) are built in for Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5. See CONTRIBUTING.md for how to update them when models change.

• This is an independent project, not made or endorsed by Anthropic.

Contributing

Improvements are welcome: see CONTRIBUTING.md. Feedback by email: [email protected].

License

MIT © 2026 Archi Dubey