Turn The Dial Down

There is a setting inside almost every model you use right now that quietly triples your bill and nobody told you it existed.

It is called reasoning effort. Some tools call it thinking budget. Some call it a mode. Whatever the label, it is a dial that decides how long the model chews on your request before it answers. Low means it fires back fast. High means it sits there and grinds, burning tokens you pay for, on a problem that might have taken twelve seconds of thought.

Here is the part that should make you sit up. You did not set that dial. Somebody at a software company set it for you, and they set it based on what makes their product look smart in a demo, not what makes your invoice reasonable at the end of the month.

I went through my own stack last week and found four places where I was paying premium prices to have a model think very hard about tasks that required roughly the intelligence of a competent intern. This is that audit, written down so you can run it yourself.

Let me give you real numbers instead of hand waving.

Artificial Analysis ran per task cost measurements on Gemini 3.8 Flash across its three effort levels at the start of September. Low came in at twenty four cents per task. Medium at forty one cents. High at fifty eight cents. That is a 2.4 times spread on the exact same model doing the exact same job.

Now look at what you get for that money. The intelligence scores across those three settings landed at 52, 57, and 59. So you pay 2.4 times more and you get a seven point bump on a hundred point scale. On some tasks that seven points is the whole ballgame. On most of the tasks you actually run every day, it is noise you would never notice.

That is the trade nobody explains when they hand you a chat box.

It gets more interesting when you look across models. GPT-6 Astra, which OpenAI shipped on September 3, runs roughly thirteen times more expensive per token than Gemini 3.8 Flash. Thirteen times. Astra is a genuinely different animal, built to run long tasks across browsers, spreadsheets, and desktop apps rather than just answer questions. When you need a model to operate software for forty minutes without losing the plot, that price is fair. When you need it to rewrite a subject line, you have just paid Michelin prices for a grilled cheese.

The dial and the model choice are two separate decisions, and most people are getting both wrong in the same direction. Up.

Watch this pattern once and you will see it everywhere.

Anthropic made Fable 5.1 generally available on September 1. In Claude Code, the developer tool, Fable 5.1 defaults to High effort. In claude.ai, the consumer chat interface, it defaults to Medium. Same model. Same company. Two different defaults, chosen for two different situations.

That is not an accident and it is not sloppiness. That is a company reasoning carefully about who is doing what and how much thinking each job deserves. A developer refactoring a codebase needs the grind. Somebody asking a question in a chat window usually does not.

The lesson is not that Anthropic picked right or wrong. The lesson is that somebody made that call for you, and the moment you run anything at volume, that call becomes your line item.

Vendors also keep moving the dial around. GLM-5.3 removed the ability to turn thinking off entirely. Fable 5.1 lets you switch effort mid conversation without blowing up your cache, which is a genuinely useful piece of engineering. Meta shipped Muse Spark 1.3 on September 2 with twenty percent fewer tool calls and twenty five percent fewer tokens than the version before it, which is a different route to the same destination, making the thinking cheaper instead of asking you to buy less of it.

All of which means the dial is not settling down. It is becoming the most important routing decision in your stack, more important than which model you picked, because the same model at two effort levels is functionally two different products at two different prices.

Go pull up every place you use AI at any kind of volume. API calls, automations, scheduled jobs, the stuff running inside your Make scenarios, the assistants your team pings forty times a day. For each one, answer four questions.

One. What is the failure cost if this is slightly worse?

Not catastrophically worse. Slightly. A subject line that is a six instead of a seven. A meeting summary that misses a minor point. A first draft that needs one more pass.

If the answer is "almost nothing, I review it anyway," that workflow does not need high effort. It needs low effort and your eyeballs.

If the answer is "it goes straight to a client without me seeing it," now we are somewhere else. Turn it up, and while you are at it, ask yourself why that is running unreviewed.

Two. How many times a day does this run?

This is the multiplier that turns a rounding error into a real number. A workflow you trigger twice a week can run at max effort forever and you will never feel it. A workflow that fires two hundred times a day at fifty eight cents instead of twenty four cents is costing you an extra sixty eight dollars a day. That is twenty five thousand dollars a year on a dial position.

Volume is where the money is. Always has been.

Three. Is the hard part the thinking or the context?

This one separates the people who understand these tools from the people who are guessing.

A lot of tasks feel hard to the model not because the reasoning is difficult but because you gave it nothing to work with. You asked it to write a proposal without telling it what the client said on the call. So it thinks harder, burns more tokens, and produces something generically competent that you then rewrite.

Feeding the model better context is almost always cheaper than making it think harder. Better context at low effort beats no context at high effort, most days, on most jobs. If your instinct when output is bad is to reach for the dial, check the input first.

Four. Does the output get checked by a human before it matters?

If yes, run it low. You are the quality layer. That is what you are for.

If no, either turn the dial up or add a human. Those are the two options. There is not a third one where you run unreviewed output at low effort and hope.

Once you have answers, write them down. Not in your head. In a table you can hand to somebody else.

Here is the shape I use. Four columns. Workflow. Volume per day. Effort level. Why.

Then three tiers.

Tier one, low effort. Extraction, formatting, classification, summarizing something you are about to read anyway, first drafts you will rewrite, tagging, routing, sorting. Anything where you are using the model as a fast pair of hands rather than a brain. This is most of what you do. Honestly. If less than half your workflows land here, you either have an unusual business or you have not been honest in the audit.

Tier two, medium effort. Analysis you will act on. Client facing drafts that get one review pass. Research synthesis. Anything where a subtle miss costs you real time downstream. This is the default for work that leaves your desk.

Tier three, high effort. Multi step agent runs. Code that touches production. Financial modeling. Anything running unattended for more than a few minutes. Anything where being wrong is expensive and you will not catch it.

The discipline is not building the table. Anybody can build a table. The discipline is defaulting everything to tier one and forcing yourself to justify every promotion. Most stacks are built the other way around, everything starts at high and nothing ever comes down, because nobody has a reason to turn it down and everybody has a vague fear of turning it down.

Give yourself the reason. It is money.

FROM THE AI NEWSROOM

The exact systems behind everything in this issue. Routing tables, gate logic, the scenario templates, and the review cadences, built out step by step so you can copy them straight into your own stack. One time, forty seven dollars.

Get the Blueprint for $47

.....

You are going to have a workflow where you genuinely do not know if low effort is good enough. Here is how to find out in about twenty minutes instead of arguing about it for a month.

Pull twenty real inputs from that workflow. Not made up examples. Real ones, from the last two weeks, including the messy ones.

Run all twenty at your current effort level. Save the outputs. Run the same twenty at one level lower. Save those.

Now strip the labels. Shuffle them. Grade all forty against whatever standard actually matters for that job, and grade them cold, without knowing which is which. Score them one to five.

Then unmask and compare averages.

What you will find, most of the time, is that the gap is smaller than you assumed and concentrated in a specific kind of input. Maybe low effort handles nineteen of twenty fine and falls apart on the one with the weird edge case. That is not an argument for running everything high. That is an argument for a routing rule that catches the edge case.

If the two sets score within a couple tenths of each other, you have your answer. Turn it down and go do something else with the money.

Run this in Galaxy.ai if you want to test the same prompt across several models and effort levels without holding four subscriptions. That is genuinely the fastest way to do a head to head, and you can find it at galaxy.ai. It is what I use when I am deciding where a workflow belongs.

A table nobody follows is decoration. Here is how you make it structural.

If you are running workflows through Make.com, build a router module at the front of your AI scenarios. One filter checks the job type coming in. Extraction and formatting go down one branch to a cheap model at low effort. Analysis goes down a second branch. Anything flagged high stakes goes down a third.

The trick is making the classification explicit at the point of entry rather than baked into each module separately. When you decide to move a whole category of work down a tier, you change it in one place instead of hunting through fourteen scenarios.

Add a second thing while you are in there. Log the effort level with every run. When you look at your bill in a month and it went up forty percent, you want to know whether that is more volume or higher effort, because those are completely different problems with completely different fixes.

If you are running the automation side through Go High Level, same principle applies at the workflow level. Every AI action gets tagged with its tier. Every tier gets reviewed monthly. You can set that up at gohighlevel.com.

While the dial is the bigger lever, model choice still matters, and September gave you more options than you had in August.

Z.ai shipped GLM-5.3-Flash, the first natively multimodal GLM-5, with a million token context window and MIT licensed weights. MIT licensed matters more than most people register. It means you can run it yourself, on your own hardware, without asking permission or paying per token. For high volume, low complexity work that never needs to be frontier smart, self hosting an open weight model is the version of turning the dial down that goes all the way to zero.

That is not for everybody. If you do not already have somebody who can stand up infrastructure, the total cost of you learning to do it is higher than the API bill you are trying to avoid. But if you do have that person, and you have a workflow running tens of thousands of times a month, run the math.

Meta's Muse Spark 1.3 is the other direction worth watching. Fewer tool calls and fewer tokens for the same work is the cleanest kind of savings, because it costs you nothing in quality. When a vendor ships efficiency instead of asking you to buy less, take it.

Here is what the audit usually turns up, and it is not really about the dial.

Most of the overspend is not a settings problem. It is a workflow that should not exist running very efficiently. You are paying a premium model to think hard about a report nobody reads. You are burning high effort tokens generating variations of something you always end up rewriting from scratch.

The dial audit is worth doing on its own terms. You will find money. But the reason I do it every quarter is that it forces me to look at every AI workflow in the business and say out loud what it is for. About one in five does not survive that sentence.

Turning off a workflow saves you a hundred percent of its cost. No dial gets you that.

FROM THE AI NEWSROOM

For operators who would rather build it with someone than build it alone. The full stack, installed alongside you, with the decisions made in the room instead of guessed at. Ninety seven dollars.

Join the Accelerator for $97

.....

Do not do the whole audit. You will not finish it and you will feel bad.

Pick the one AI workflow that runs the most times per day. Just one. Find its effort setting. Run the twenty input test. If the gap is small, turn it down and put a calendar reminder two weeks out to check whether anything broke.

That is it. One workflow, twenty inputs, one dial.

If the answer is what it usually is, you will have found real money in an afternoon and you will have a reason to do the next one. And if the answer goes the other way, you will have learned something worth more than the money, which is that the workflow genuinely needs the horsepower and you should stop second guessing it.

The dial is not going away. It is becoming the main thing you tune. Might as well get good at it now, while most people still do not know it is there.

Jordan

The AI Newsroom | Practical AI for people with a business to run.