The cheapest tier in every major AI lineup is now the one built for document work, and this article answers the question most owners get backwards: when is the cheap model the right model, and how do you prove it on your own paperwork?
You have a stack of invoices and a sense that AI should be handling them by now.
So you open the tool, and the first decision you hit is a menu. Five or six model names, none of which mean anything, and a price difference between the top and the bottom that runs into the double digits. You pick the smartest-sounding one, because it is your business and this seems like the wrong place to be cheap.
That instinct is costing you money, and it is also making your results worse than they need to be.
Here is the direct answer. For reading invoices, receipts, intake forms, and anything else with a fixed set of fields to pull out, you should be using the cheapest tier your vendor offers, not the most capable one. The vendors have now explicitly labeled that tier for this work. Anthropic's own model documentation describes Claude Haiku 5.5 as "for high-volume, latency-sensitive tasks such as classification, extraction, and routing." Google's Gemini API docs describe Flash-Lite as "optimized for high-volume agentic tasks, translation, and simple data processing." And when OpenAI shipped GPT-6.1 Terra on October 26, it priced it at $1 per million input tokens and $5 per million output, which the company states is 50% below both of its previous comparable tiers. The example OpenAI leads with in its own announcement is reading an incoming invoice and entering the details into accounting software.
The frontier spent this month moving toward your filing cabinet, and it moved there cheap.
But there is a second half to that answer, and it is the half that decides whether this works for you. The cheap tier has a floor, it is jagged, and the single biggest factor in your results is not which model you pick at all. It is how you hand the document over.
I am going to show you the research on both, then give you the test to run on your own paperwork this week.
Key Takeaways
- On clean invoices, a published Fraunhofer IAIS benchmark found the cheap tier scoring 95.70% against the flagship's 96.50%, a gap of eight tenths of a percentage point across a price difference of many multiples.
- The handoff matters more than the model. In that same study, feeding the document image directly scored 87.46% on receipts while converting it to text first scored at most 47.00%.
- The cheap tier has a real floor and it is unpredictable. One small model dropped to 53.92% on receipts and a 4-billion-parameter open model scored 45.69% on clean invoices.
- Your costs hide in exceptions, not keystrokes. Ardent Partners put best-in-class accounts payable exception rates at 9.0% against 22.0% for everyone else.
- The constraint on getting value from this is no longer capability or price. It is whether you have defined one task precisely enough to hand it off.
The problem is not the model. It is the menu.
I watch business owners do this every week in our community, and I have done it myself.
You know AI should be handling your repetitive document work. You have read that it can. And then you sit down to actually start, and you run into a decision you have no basis for making. Which model. Which tier. Which reasoning setting. Nobody publishes a plain-English map of which one is for what.
So you do the sensible thing a careful person does. You pick the most expensive option, because the downside of a mistake on your books feels worse than the cost of overpaying.
Here is how common that is. The day GPT-6.1 Terra launched, the biggest discussion thread about it on Hacker News was not about capability. It was about confusion. One developer wrote: "I can't tell you the difference between the 5-6 different models put out by the Anthropics/OpenAIs of the world. At this point I just try to track the flagship and then the cheapest one." Another said the thing I hear from owners constantly: "Or I'm going to give up and not use any."
Those are professional software developers. If they have given up on tracking the lineup, you are allowed to.
And the paralysis has a real cost that compounds. Census Bureau data from May 2026 put national business AI use at 19.8%, with 37% of firms over 250 employees using it against under 20% of firms with four or fewer. The more telling part of that release: use rose among firms with 20 or more employees and showed no significant change for firms under 20.
The small end of the market did not get locked out by price. Access got cheaper for everyone at the same time. What the big companies have is not a better model. It is a written description of how their work gets done, usually because a regulator made them write it.
That is the real gap, and it is fixable without buying anything. But I want to put the evidence in front of you first, because the next part is counterintuitive enough that you should not take it on my word.
What the research actually found
In August 2025, a team at Fraunhofer IAIS and the Lamarr Institute published a benchmark on invoice processing, testing eight models across three open datasets: 500 synthetic invoices, 1,000 scanned receipts, and 350 scanned invoices. It is the closest thing to a direct answer for this specific job that exists in the public literature.
Three findings matter for you.
First, the cheap tier is barely behind on clean work. On clean invoices, Gemini 2.5 Flash-Lite scored 95.70% against Gemini 2.5 Pro at 96.50%. On scanned invoices, 91.64% against 92.71%. On scanned receipts, Flash-Lite at 85.34% actually beat the mid-tier Flash at 82.15%. In the OpenAI family, GPT-5-mini beat GPT-5-chat on clean invoices.
You are not paying a large multiple for an extra percentage point. You are paying it out of caution that the data does not support on this kind of task.
Second, the floor is real and it is not where you would guess. The same study had GPT-5-nano drop to 80.61% on scanned invoices against GPT-5-chat's 88.01%. GPT-5-mini collapsed to 53.92% on receipts. An open 4-billion-parameter model scored 45.69% on clean invoices, and the authors name this directly as a capability threshold below which smaller models stop being effective.
So "cheap" is not a reliable label. The second-cheapest tier usually holds. The bottom of the lineup does not, and it fails without warning you.
Third, and this is the finding I want you to remember: the pipeline beat the model. Feeding the document image directly to the model scored 87.46% on receipts. Converting the document to markdown text first, then handing that text over, scored at most 47.00%. On scanned invoices the comparison was 92.71% against 64.03%. And on clean invoices, the conversion step flattened every single model into the same 84 to 85% band. The authors write that the conversion, "rather than the LLMs' reasoning abilities, becomes the primary limiting factor."
Forty percentage points, on a step that happens before the model ever sees the work.
This is not a new idea so much as a confirmed one. Back in 2023, a Stanford team published FrugalGPT, observing that LLM API fees differed by two orders of magnitude and showing that a routing cascade could match the best individual model's performance with up to 98% cost reduction on their test datasets. That price spread has not closed. It has just been labeled. OpenAI's own current price list runs from $0.10 to $12.50 per million input tokens. Anthropic's runs from $0.10 for Haiku 5.5 to $4 for Opus 5.5.
One more piece of evidence, because it cuts against the easy version of this story. The largest peer-reviewed field experiment on AI at work, Dell'Acqua and colleagues with 758 BCG consultants, now published in Organization Science, found a 12.2% increase in tasks completed and 25.1% faster work on tasks inside AI's competence. On tasks outside it, the control group was right 84.5% of the time while the AI-assisted groups were right 60% and 70.6%. And the arm that received prompt engineering training did worse.
Read that honestly. Better prompting did not rescue a badly chosen task. Choosing the right task did.
The system: route the work, do not shop for models
Here is how I think about this, and it is a routing problem, not a purchasing problem.
Stop asking "which model is best." Start with two tiers and a written rule.
Your default tier is the second-cheapest model your vendor offers. Not the cheapest, because of that floor. Everything with a fixed output goes here: pulling fields out of documents, sorting incoming email into categories, formatting, routing, first-pass summarizing. This is the extraction tier and the vendors now say so themselves.
Your fallback tier is one good premium model. Anything where the output feeds a judgment you make, where the work is open-ended, or where the stakes of a quiet mistake are high. This is the judgment tier.
Then write the escalation condition down. Not "use your judgment." Something like: scanned documents with handwriting go to the premium tier, anything over a certain dollar amount gets a human check regardless of tier, and anything the cheap tier marks as unclear gets escalated rather than guessed.
And fix the pipeline before you touch the tier, because the research says that is where your forty points are hiding. Give the document to the model as a document. Stop converting to text first. Stop screenshotting a PDF. Every conversion step you remove is accuracy you get back for free.
Two more pieces make this a system rather than a tip.
Define the output, including the behavior when something is missing. An invented value is worse than a blank one, because a blank one gets caught. Specify your fields, your formats, and an explicit instruction that anything not clearly present gets marked unknown rather than inferred.
Measure exceptions, not accuracy. This is the part almost everyone skips. Ardent Partners' 2025 accounts payable benchmark put the average all-in cost per invoice at $9.40, with best-in-class at $2.78 against $12.88 for everyone else. But the number underneath that is the exception rate: 14% overall, 9.0% best-in-class, 22.0% for everyone else, and exceptions ranked as the number one AP challenge at 53% of respondents. Their sample skews large, with 53% of respondents over $1 billion in revenue, so treat the dollar figures as directional for a small business. APQC's open benchmarking puts the median at $6.00 per invoice across 5,846 organizations, which tells you the same thing with a different population.
The lesson transfers regardless of your size. The cost is not in the documents that go smoothly. It is in the ones that do not fit, and those are exactly the ones a cheap model will hand back wrong if you have not told it what to do with them.
This is why the people getting real results are narrowing scope rather than expanding it. The practitioner community has moved the same direction: this week's trending posts among developers include cutting an agent's tool list from fifteen to four purely to reduce cost, and an agentic development environment built so agents never touch the main working copy. Containment, not capability.
Seven steps to have this running by Friday
1. Inventory the repetitive document work, then pick exactly one task. High volume, fixed output fields, and recoverable if it is wrong once. Not the task you hate most. The task that is most defined.
2. Narrate it out loud while you do it, exceptions included, and have AI turn the transcript into a procedure. This is the step people skip and the step that makes everything after it work. Use this prompt:
[The Job]
Turn my spoken walkthrough of this task into a clean written procedure.
This is for: handing to an AI and to a person.
It matters because: the task currently exists only as something I do without thinking.
[The Background]
Here is what you need to know: here is the transcript of me doing the task out loud [PASTE IT].
Do not use: anything you infer that I did not say; list your questions at the end instead.
[The Deliverable]
Return: a numbered procedure with inputs, steps, decision points, exceptions, and a definition of a correct result.
Must include: a separate list of every place my narration was ambiguous or skipped a step.
Optimize for: completeness over polish.
[The Questions]
Ask me any questions you have.
3. Shorten the handoff. Map every step between the document arriving and the model seeing it, then delete the conversions. Feed the document as a document. This single change was worth forty points in the research and it costs you nothing.
4. Specify the output and ban guessing. Field names, formats, and an explicit rule that anything unclear comes back marked unknown. Paste that specification at the top of every run.
5. Build a twenty-document test set out of your own paperwork, including five deliberately awful ones. Phone photos, skewed scans, the vendor whose invoice format makes no sense. A benchmark on somebody else's documents tells you nothing about yours.
6. Score the cheap tier and the premium tier on the same twenty, field by field. Count invented values separately from blanks, because those are different failures. Note whether errors cluster in one field or one document type, since clustered errors are fixable and random ones are not.
7. Write the decision down with a date and a review date. Keep, fix one thing and retest, or kill it. Then do not start a second task until this one has run clean for a month.
That last line is the one I would underline. Census research published in April 2026 found that while 18% of firms use AI in a business function, 57% of adopters use it in three functions or fewer, and breadth of task integration correlated with commercial performance. The firms getting results are not the ones with the most tools. They are the ones who finished something.
Frequently Asked Questions
Won't a cheaper model make more mistakes?
On narrow, well-specified extraction, barely. The Fraunhofer benchmark put the gap at eight tenths to one and a tenth percentage points against the flagship. But the very cheapest tier is erratic, dropping to 53.92% on noisier documents in the same study. Test on your own paperwork rather than trusting the tier label.
Should I convert my PDFs to text before feeding them in?
No. This is the most common mistake and the most expensive one. Converting documents to markdown first cut accuracy from 87.46% to 47.00% on receipts in the Fraunhofer study, and flattened every model tested into the same narrow band on clean invoices. Hand over the document itself.
How do I know whether this is actually saving me money?
Benchmark against APQC's median of $6.00 per invoice across 5,846 organizations, or Ardent Partners' $9.40 average and $2.78 best-in-class. Then track your exception rate, because that is where the cost hides. Best-in-class exception rates were 9.0% against 22.0% for everyone else.
Will better prompting fix a disappointing result?
Not reliably. In the Organization Science field experiment with 758 consultants, the group given prompt engineering training performed worse than the group that was not, and AI assistance reduced accuracy by roughly 19 percentage points on tasks outside its competence. Pick a better task before you write a better prompt.
Is everyone else already doing this?
No, and this is the part people get wrong. Census data from May 2026 put national business AI use at 19.8%, under 20% for firms with four or fewer employees, with no significant increase among firms under twenty. You are not late. You are early, and the tools just got cheap.
The cheapest thing in the lineup is pointed at your filing cabinet
Go back to where you started. A stack of invoices, a menu you cannot read, and a reasonable instinct to be careful.
That instinct was right about caution and wrong about where to apply it. Being careful does not mean paying for the biggest model. It means testing on your own documents, counting your exceptions, and writing down what you learned so the next sales pitch does not overwrite it.
The frontier came for the boring work this month, and it came in at a dollar per million words of input. The constraint is not the price anymore. It was never really the price.
The constraint is that scaling this requires knowing exactly how your work gets done, step by step, including the exceptions. The big companies wrote that down because someone made them. You have it in your head.
That is fixable in an afternoon, and it does not require buying anything. Pick one task. Narrate it out loud. Hand the transcript to AI and let it write the procedure. Then hand the procedure to the cheap model and test it on twenty of your own documents.
You did not need a better model. You needed to write down what you already know.
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 for entrepreneurs using AI, the AI Insiders membership, and the Perfect Prompt Framework used by thousands of business owners to get usable output from AI on the first try. He is a regular speaker on practical AI for small business. He has also personally overpaid for a premium model to do data entry for longer than he would like to admit in print.
