Our team will be out of office on Friday, May 1, 2026. We’ll be back and ready to assist you starting Monday, May 4th.

Which AI model should I actually be using for everyday business tasks?

Contents

The cheapest one that clears your quality bar, not the most capable one you can afford. This article shows you how to find that model in about 20 minutes, using your own work as the test.


The Hook

Anthropic published two numbers yesterday that, read together, tell you something nobody in the coverage bothered to point out.

Claude Haiku 5.5 costs 90 percent less per token than its predecessor on tasks under 100,000 tokens. And in the same post, the company notes that tasks under 100,000 tokens made up around 90 percent of all requests to the previous Haiku model.

Ninety percent cheaper on the work that was ninety percent of the work.

That is not a discount on an edge case. That is a discount on almost everything you do, and most business owners will never collect it, because they made a model choice once and turned it into a default they have never revisited.

So here is the direct answer to the question in the headline. For everyday business tasks, you should be using the cheapest model that produces an acceptable result, not the best result, and you should determine which one that is by running ten of your own real inputs through it rather than reading somebody's benchmark. The rest of this article is how to do that and why the instinct to upgrade instead is costing you money.

Key Takeaways

  • Price cuts at the bottom of the market are not an upgrade budget, they are a reason to stop reaching for the most capable model by default.
  • Stanford HAI's AI Index tracked the price of GPT-3.5 level performance falling from $20.00 per million tokens in November 2022 to $0.07 by October 2024.
  • Your bill is driven by which tasks you route to a reasoning model, not by the sticker price, because reasoning models can consume far more tokens internally than they show you.
  • Model choice is becoming a per-task setting rather than a purchase, which makes it a process decision you own rather than a vendor decision you accept.
  • A 20 minute audit of your three highest-volume tasks will usually find real savings without changing a single output you rely on.

The Problem

Let me describe the actual situation most small businesses are in, because the articles about this never do.

You picked an AI tool. You picked a plan. Somewhere in a dropdown there is a model selector that you set once, probably to whatever was described as the best one, and you have not touched it since. You have no idea what any individual task costs you. You see one number on one bill at the end of the month and you either wince or you do not.

That is not negligence. That is what happens when the pricing is per token and your work is per task and nobody has ever translated between the two for you.

And it got worse this year, not better, because the number of options multiplied. There is a cheap tier and an expensive tier and now an adjustable effort setting inside a single model, plus routers that promise to pick for you, plus open-weight models you could run yourself. More choice did not make this easier. It made the default more expensive to keep.

Here is what the default is costing you. If you are sending classification, tagging, summarizing, and first-pass drafting to a model built for hard reasoning, you are paying frontier prices for pattern work. Not a little more. Often an order of magnitude more.

But what if the fix is not switching providers, and not buying a router, and not learning anything technical at all.

The Evidence

Five data points, and the fourth one is the one almost nobody accounts for.

The floor dropped further and faster than the ceiling. Stanford HAI's AI Index tracked the price of achieving GPT-3.5 level performance from $20.00 per million tokens in November 2022 down to $0.07 by October 2024. Epoch AI's work on the cost to reach a fixed capability bar is cited at a median of roughly 50 times cheaper per year, with some task categories far steeper. Those are research figures, not vendor marketing. And they describe a specific thing: holding capability constant, the price collapses. The newest frontier model does not get cheaper nearly as fast.

Yesterday's cuts were concentrated at the bottom. From Anthropic's own announcement: Haiku 5.5 costs around 75 percent less to run on average than Haiku 4.5, 90 percent less per token on tasks under 100,000 tokens, and half the price on longer ones. Cache reads on Sonnet 5.5 were halved, which the company puts at roughly 20 percent cheaper on most long-running work. OpenAI moved in the same direction, with GPT-6.1 Sol reported at $2 per million input and $10 per million output tokens, described by one roundup as a fifth of the prior flagship's standard pricing. I will flag that honestly: the OpenAI figures come from third-party roundups and the model naming is inconsistent across sources, so verify before quoting.

The open-weight community is pushing from underneath. The loudest open model discussion I saw this week was about a 27 billion parameter model claiming 96 percent of a much larger model's performance at roughly one seventh the size. Separately, a llama.cpp change adding a GPU cache for mixture-of-experts models kept in host memory drew heavy engagement for one reason: it helps people whose models do not fit in their graphics card. That is not a frontier-capability conversation. That is a running-this-affordably conversation, and it is where the practitioner energy is.

Cheap per token does not mean cheap per task. This is the one that will actually explain your bill. One cost analysis warns that reasoning models can consume up to 100 times more tokens internally than they output. You see three paragraphs. You assume you paid for three paragraphs. The deliberation happened out of sight. Which means your spend is not a function of which model you picked, it is a function of which of your tasks you asked a model to think hard about. And the average market price confirms the gap between the cheapest tier and reality: Silicon Data's index put the average at roughly $0.97 per million tokens on September 1, 2026, while Google's Gemini Flash-Lite 3.1 lists at $0.25 per million input tokens. The cheap tiers exist. Most usage is not in them.

The productivity upside is real but it is not about the model. The strongest study I found is from the University of St Andrews, using the UK Small Business Survey across just under 10,000 small and medium firms, suggesting AI adoption produces productivity gains between 27 and 133 percent. The authors flag a serious caveat themselves, which is that the lowest-productivity firms are the most likely to adopt, so selection effects matter. The OECD's estimate is more conservative, at 0.2 to 1.3 percentage points of annual labour productivity growth across G7 economies over a decade. Business.com's 2026 outlook reports an average of 5.6 hours per week saved per small business worker, which is self-reported survey data.

Notice what is missing from every one of those studies. None of them measures which model you used. The gains come from doing the work differently, not from buying the better engine.

The Solution

Here is how I think about this now, and it changed when I stopped treating model choice as a purchase and started treating it as a routing decision.

You have a floor model and a ceiling model. The floor is the cheapest thing that clears your bar. The ceiling is the expensive thing you escalate to when the floor fails.

Most people run the ceiling by default and never discover the floor was fine. That is the entire mistake, and it is expensive precisely because it never produces a visible problem. You get good results. You just overpay for them every single time.

Run the floor by default and escalate on failure. That is the whole architecture.

And three things that shipped this week make it easier than it has ever been to do.

First, Haiku 5.5 is the first Haiku with an adjustable effort setting. Read what that actually means for you. The decision moved out of the model dropdown and into a per-task setting. Which makes it a process decision you control rather than a product decision you buy once.

Second, the small models got genuinely good at being sub-agents. Anthropic's own framing for Haiku 5.5 includes pairing it with larger models as a sub-agent on coding work. That is a two-stage pattern you can copy for almost anything: let the cheap model gather, sort, extract, and format, and let the expensive model do only the judgment at the end.

Third, cache read prices came down on the mid-tier model, which specifically rewards the pattern where you work from the same context repeatedly. If you have been avoiding long-running work on cost grounds, that math moved.

Candidly, the hardest part of this is psychological, not technical. There is something in us that reads "cheaper" as permission to buy more, and reads "the best model" as the responsible choice. On high-volume routine work, the best model is the wrong choice, and paying for it is not responsibility. It is just a habit nobody has ever asked you to examine.

Practical Steps

Twenty minutes. Five steps. You will not switch platforms and you will not lose any result you currently depend on.

Step 1. Minutes one through five. List what you run. Write down every AI task you perform in a typical week with an honest guess at how many times you do it. Do not organize by importance. Organize by volume, because volume is what drives cost. Flag anything you do more than ten times a week.

Step 2. Minutes six through ten. Mark the three that are probably overpaying. Go through your list and separate the tasks that genuinely require reasoning from the ones that are pattern work. Classification, tagging, summarizing, extracting, reformatting, and first-pass drafting are almost always pattern work. Then pick the three highest-volume pattern-work tasks. Those are your candidates.

Step 3. Minutes eleven through fifteen. Test one for real. Take your top candidate. Grab five real inputs you have actually used. Run them through the cheaper model. Compare against what you get now. Your bar is acceptable, not identical. Write down the one difference that would mean do not switch, and check whether it appeared.

This step is the whole audit. Everything before it is preparation and everything after it is bookkeeping.

Step 4. Minutes sixteen through eighteen. Write the rule down in one paragraph. Your default model, your escalation trigger, and the one task type that always gets the expensive option regardless. One paragraph. Paste it somewhere your team can see it. A change you do not write down reverts inside a week, and I say that from experience.

Step 5. Minutes nineteen and twenty. Set one tripwire. A single spend alert, at a threshold that would genuinely concern you, on the platform where most of your usage lives. One. Not a dashboard. If you set five alerts you will mute all of them.

Then two follow-ups that are worth more than the audit itself.

Shorten your inputs before you shrink your model. Look at exactly what you send, including attached files and standing instructions. Most people attach an entire document when the answer depends on two sections. Attached context is the fastest way to turn a cheap task into an expensive one, and trimming it costs you nothing in quality.

Recheck in 90 days. Prices and small-model capability are both moving fast enough that a decision made today will be partly wrong by January. Put a 20 minute recurring block on the calendar with three things to check. You do not need to follow AI news. You need to re-run your own test twice a year.

Here is the prompt I would start with, built on the Perfect Prompt Framework 2.0.

[The Job] Identify the cheapest, smallest model that produces an acceptable result for [NAME A TASK], not the best result. This is for: [ME, WHO HAS BEEN OPTIMIZING FOR THE WRONG THING]. It matters because: acceptable and repeatable beats excellent and expensive on work I run a hundred times a week.

[The Background] Here is what you need to know: the task is [DESCRIBE IT]. Acceptable means [DEFINE IT PRECISELY]. Unacceptable means [DEFINE THAT TOO]. My volume is [HOW OFTEN]. Do not use: any recommendation based on which model is most capable.

[The Deliverable] Return: a recommended starting model and the specific test that confirms it clears my bar. Must include: what I lose by using it and whether that loss matters at this volume. Optimize for: cost per acceptable result.

[The Questions] Ask me any questions you have.

Frequently Asked Questions

Will using a cheaper AI model make my work worse?

On routine pattern work, usually not in any way your customer would notice. On tasks requiring multi-step reasoning or nuanced judgment, often yes. The point of testing five of your own real inputs is that it answers this for your specific work rather than in the abstract.

Why is my AI bill higher than my usage seems to justify?

Most likely because some of your tasks are being handled by a reasoning model that consumes far more tokens internally than it shows in the output. One cost analysis puts that gap at up to 100 times. Attached files and long running conversations are the other two common causes.

Should I buy a model router to handle this for me?

Probably not first. A router solves the selection problem, but selection is not usually where the money goes. Write your own routing rule for three task types, shorten your inputs, and set a spend cap. If you still have a cost problem after that, then consider tooling.

How often should I revisit my model choice?

Quarterly is enough for most small businesses. Capability at the cheap end has been improving fast and prices have been falling, which means last quarter's correct answer gradually becomes an overpayment rather than suddenly breaking.

Does running an open-weight model myself save money?

Sometimes, but the savings are rarely the real reason to do it. Self-hosting trades a predictable bill for hardware, setup, and maintenance you now own. For most small businesses the better first move is routing more work to a cheap hosted tier.

The Close

Ninety percent cheaper on ninety percent of the work. Those two numbers from one announcement post are the clearest picture anyone has given us of where this is heading.

Intelligence got cheap. That is now a settled fact, and the research backs it: holding capability constant, the price has been collapsing by roughly an order of magnitude a year.

Judgment about where to spend it did not get cheap. It did not get anything. It is the same work it always was, and it is the work almost nobody does, because it is unglamorous and nobody is selling it to you.

So do the unglamorous thing. Take your three highest-volume AI tasks and find out, with your own inputs, what the cheapest acceptable answer is. Write the rule in one paragraph. Set one alert.

Twenty minutes. And then every single time you run those tasks for the rest of the year, you are collecting a discount that was published in a blog post you never read.

The best model is not the responsible choice. The cheapest one that clears your bar is.


About the author. Jonathan Mast is the founder of White Beard Strategies, where he teaches non-technical business owners how to use AI to amplify the skill and experience they already have. He runs a 500,000 member Facebook community, the AI Insiders membership, and ongoing live trainings built on the Perfect Prompt Framework 2.0. He has rebuilt a business from zero more than once, which is a reliable way to develop opinions about spending money on things you do not need.

About the Author