Overkill AI Check is a work in progress. The public calculator works today, but I’m still refining the scoring and usage estimates. ChatGPT support, automatic routing, and calibration from real usage are next steps, not finished features.

I kept seeing the same thing. People were running Claude Opus 5.5 on high effort, sometimes even Fable on high, to summarize a PDF or clean up an email. These were simple reading and writing tasks.

It works, and the answer comes out fine. But a lighter model can often do that job well enough, without using so much of your allowance.

So I built Overkill AI Check to answer one question before you hit send: does this task actually need the big model?

The problem is hitting the limit

Most people I built this for use Claude through a subscription like Pro, Max or Ultra. There’s usually no separate charge for each message within the included allowance, so it feels free. But the allowance has an end.

Pro and Max have five-hour session limits as well as weekly limits. Verified Ultra accounts have no five-hour limit, but still have a weekly one. Fable has its own restrictions: on Pro it uses separately purchased usage credits; on Max and Ultra it can use up to half the weekly allowance. Those are the terms on Claude’s pricing page as of November 8, 2026, and they can change.

Model choice, effort, long conversations, large attachments and tool use all affect consumption. Anthropic lists those factors in its usage guidance.

If you run everything on the biggest model, you can hit the wall in the middle of the afternoon. That might be the exact moment you need Opus for something that’s actually hard.

Two dials to adjust

There are two settings worth thinking about. The first is the model. In the tool’s current lineup, Haiku is the lightest, followed by Sonnet and Opus, with Fable at the top. The second is effort, which controls how much reasoning the model puts into an answer.

It’s easy to set both high and leave them there. But straightforward reading, extraction, summarizing and rewriting don’t always benefit enough to justify that choice. I wanted a quick way to stop and make the choice deliberately.

What Overkill AI Check does

You describe what you’re doing: your prompt, what you’re trying to get done, how much material Claude has to read, how much reasoning it takes, what a wrong answer would cost you, and which connectors you’re using. The tool scores those inputs and recommends a model and effort level to start with.

Then it compares estimated usage with Opus on high effort and Fable. For example, the contract preset asks for dates and payment terms from a 20-page document. The current rules recommend Haiku at low effort, and the calculator estimates that Opus on high would use dozens of times more allowance under its assumptions.

That’s a modeled comparison, not a measured benchmark or a promise that both answers will be identical. Extracting dates is also different from interpreting a contract’s legal risks. Change the task, reasoning or stakes and the recommendation can change too.

It works in the other direction. The multi-file code migration preset recommends Opus at high effort. A difficult investigation can also justify a stronger model. I want the tool to make that call plainly when the task warrants it.

Why it doesn’t use AI to decide

I made this choice on purpose. A tool meant to save your usage shouldn’t spend your usage just to give you a recommendation.

The recommendation comes from a fixed set of rules. The same selected inputs get the same result, and you can open the scoring breakdown to see why. The “Suggest fields from my text” button uses a rough keyword match; it doesn’t send your prompt to a model. You still need to check its picks.

There’s no login and no API key. Your prompt and goal stay in your browser and aren’t saved. The calculator remembers only your selected plan between visits.

Built for people who don’t live in AI

Plenty of people are technical without knowing what a token is, what effort does, or how plan limits work. So every field has a small “?” that explains it in plain language.

I also built an internal version for my team at TechInSF. Instead of filling in every detail, you pick your division and the job you’re doing, like “troubleshoot a client IT issue,” “write a proposal,” or “triage a security alert,” and the details fill in for you. I put that list together from what each team actually does day to day.

The public version starts with example tasks instead. You can try a contract summary, a client email, a code migration or an overnight research agent, then change the details to match your own work.

Being honest about the numbers

The usage guidance I checked explains what affects your allowance, but it doesn’t provide a formula that turns an arbitrary task into an exact percentage of your subscription limit.

The calculator fills that gap with assumptions about session size, token counts, repeated context, effort and relative API prices. API prices are a proxy in this model; they don’t establish how Anthropic deducts subscription usage. Both the percentages and the “times more usage” comparisons are estimates.

That’s a big reason this is still a work in progress. The scoring hasn’t established that a recommended model will handle every task in its band, and the usage numbers haven’t been validated across a representative set of real tasks.

The next step for the team version is calibration: recording the usage percentage before and after a task, then comparing those readings with the estimate. Those readings should help improve the model, but concurrent work, rounded percentages and different task sizes make them imperfect measurements too. The public calculator doesn’t currently collect those readings.

What’s next

  • ChatGPT models, so the tool can help across more than one assistant.
  • A Claude Code plugin that routes requests to a suitable model automatically, instead of just advising.
  • Calibration from real usage, to test and improve the estimates.

Those are development plans. For now, try Overkill AI Check, look at the scoring breakdown, and treat the recommendation as a starting point. Check the output before making a lighter model your default.

Overkill AI Check is an independent tool and isn’t affiliated with Anthropic.