Choosing a large language model (LLM), the AI behind ChatGPT and Claude, means matching a model's strengths to the job, your monthly budget and your rules about where data can go. For most small businesses that means a middle-tier model such as Claude Sonnet 5.5, GPT-6.1 Sol or Gemini 3.8 Flash, plus a cheaper one for high-volume sorting.
A test score shows how a model does on a standard exam, which says little about how it will handle your invoices or your customer emails. This guide covers the decisions that matter when you pay for AI by how much you use it, and what we picked for the 15 AI agents on our own team. We checked every model and price below against the vendors' own pages on November 16, 2026.
Why the 'best LLM' question is the wrong starting point
The model is one part of a working AI setup. The information it reads, the systems it connects to and the way someone checks its work shape the result too, and a stronger model cannot make up for weak work in those places.
A better first question is what exactly you need the model to do.
An assistant that answers customer questions has different needs from one that sorts incoming invoices, and summarizing long documents has different limits from drafting sales emails. Starting with the job instead of the model avoids the most expensive mistake, which is paying for the most powerful AI while skimping on everything around it.
How to choose an LLM by use case
Anthropic, OpenAI and Google each sell their models in roughly three price tiers. Open models, whose files anyone can download and run on their own computers, cover the cases where data cannot leave your building. The table matches those tiers to common business jobs, using the models current in November 2026.
| Job | Tier | Current examples | Why |
|---|---|---|---|
| Complex reasoning, research, long multi-step work, writing software | Top tier | Claude Fable 5.2, Claude Opus 5.5, GPT-6 Astra, Gemini 4 Carbon, Gemini 4 Argon | Work where accuracy matters more than speed or cost |
| Customer support, drafting, summarizing | Middle tier | Claude Sonnet 5.5, GPT-6.1 Sol and Terra, Gemini 4 Hydrogen, Gemini 3.8 Flash | Solid quality at a fraction of the top-tier price |
| Sorting, tagging, pulling out simple details | Fast and cheap | Claude Haiku 5.5, GPT-6 Luna, Gemini 3.5 Flash-Lite | High-volume work where speed and cost per request matter most |
| Data that must stay on your own computers | Open model | gpt-oss-120b, Qwen3.8, Mistral Small 4 | Nothing leaves your building |
| A narrow, high-volume, cost-sensitive job | A small open model trained on your examples | A further-trained Qwen3.8-27B or Mistral Small 4 | A predictable job where a specialist model can beat a general one |
The vendors describe their models the same way. On its pricing page, Anthropic describes Claude Haiku 5.5 as built for high-volume, latency-sensitive tasks such as classification, extraction and routing. OpenAI's model guide calls GPT-6 Astra its most capable model "for the most demanding work," GPT-6.1 Sol "near-Astra performance for complex work at a lower cost" and GPT-6 Luna its most efficient model "for focused, high-volume tasks." In late October OpenAI added GPT-6.1 Terra at half Sol's price, "a fast, efficient model for visual reasoning, precise computer use, and clear, natural writing." Google's pricing page calls Gemini 3.5 Flash-Lite "optimized for high-volume agentic tasks, translation, and simple data processing." On November 10 Google released Gemini 4 Carbon, its "most capable reasoning model," and now calls Gemini 4 Argon its "efficient Gemini 4 frontier reasoning model." The two cost the same, and both prices double on December 1, 2026, when an introductory rate ends. Two days later Google added Gemini 4 Hydrogen at $1.50 to read and $7.50 to write per million, offering "speed and efficiency for everyday coding and agentic workflows."
All three open models publish their files for anyone to download. OpenAI calls gpt-oss-120b its most powerful open model and releases it under the Apache 2.0 license, which allows business use. Alibaba's Qwen team publishes Qwen3.8 on Hugging Face, and Mistral lists Mistral Small 4 under Apache 2.0 in its model catalog.
A common pattern is to start with a top-tier model to prove the job can be done, then move down to a middle-tier or specially trained model once the work runs well and its inputs and results are predictable.
LLM pricing: what each model costs
AI companies bill for the text a model reads and the text it writes, counted in tokens. A token is a small piece of text, and OpenAI's rule of thumb is that one is about three-quarters of a word in English. So the prices below, quoted per million, cover roughly 750,000 words. Text the model writes costs more than text it reads. These are the standard prices from the Anthropic, OpenAI and Google pricing pages, checked November 16, 2026. The full table, with cache and long-request rates and a log of every price change we find, is in our LLM API price dataset.
| Model | Text it reads, per million | Text it writes, per million |
|---|---|---|
| Claude Fable 5.2 | $8.00 | $40.00 |
| Claude Opus 5.5 | $4.00 | $20.00 |
| Claude Sonnet 5.5 | $2.00 | $10.00 |
| Claude Haiku 5.5 | $0.10 | $0.50 |
| GPT-6 Astra | $10.00 | $50.00 |
| GPT-6.1 Sol | $2.00 | $10.00 |
| GPT-6.1 Terra | $1.00 | $5.00 |
| GPT-6 Luna | $0.10 | $0.50 |
| Gemini 4 Carbon | $2.00 | $10.00 |
| Gemini 4 Argon | $2.00 | $10.00 |
| Gemini 4 Hydrogen | $1.50 | $7.50 |
| Gemini 3.8 Flash | $0.75 | $3.75 |
| Gemini 3.5 Flash-Lite | $0.30 | $2.50 |
Three of these prices are introductory. Gemini 4 Carbon and Gemini 4 Argon cost $4 and $20 from December 1, 2026, and Gemini 3.8 Flash costs $1.50 and $7.50 from January 1, 2027. Very long requests cost more on Claude Haiku 5.5 and on OpenAI's models, and the price dataset lists those rates.
The spread inside one vendor's lineup is wide. OpenAI's top model costs 100 times as much per unit of text read as its smallest one ($10 against $0.10), and Anthropic's costs 80 times as much ($8 against $0.10). Google's spread is narrower, under seven times until the Gemini 4 Carbon and Argon prices double.
Discounts move the numbers a lot. All three vendors charge half price for work you are willing to wait on, sent in batches. Text you send over and over, such as a long set of standing instructions, can be stored and costs a fraction of the normal rate the next time: $0.10 per million instead of $2 on Claude Sonnet 5.5.
Prices are also not directly comparable across vendors, because each one counts the same text differently. Anthropic notes that its Claude 4.7 and later models count the same text as about 30% more units than its earlier models did. Compare costs on your own requests, not on the price table alone.
Here is what this means in dollars a month. Take a request where the model reads 2,000 units, roughly 1,500 words, and writes 500, roughly 375 words. At 10,000 requests a month, Claude Haiku 5.5 costs about $4.50 and Claude Fable 5.2 about $360. At a million requests a month, that becomes $450 against $36,000. Our LLM cost calculator runs the same math with your own numbers for every model in the table.
The cheapest option per request is not always the cheapest overall. Say that at 10,000 requests a month, the cheaper model leaves 1,000 more answers that a person has to fix, at five minutes each. That is more than 80 hours of someone's time spent to save about $356 in AI fees. AI fees are also only one part of the total, and what an AI agent costs covers the setup, service and staff time on top of them.
Rent it or run it yourself: API, self-hosted or further trained
Which model you pick and where it runs are separate decisions, but each one limits the other.
Using the AI company's online service, its API, is the simplest option. Your software sends requests to Claude, GPT or Gemini and you pay for what you use, with no hardware to buy and nothing to maintain. It suits most businesses and is the fastest way to start. The tradeoff is that your data passes through the AI company's computers, and you are subject to its usage limits and price changes. Keeping your data inside one country usually costs extra. OpenAI adds 10% for processing in a chosen region on models released on or after March 5, 2026, and Anthropic charges 1.1 times the normal price to keep processing in the US on Claude 4.6 and later models.
Running an open model yourself means putting gpt-oss-120b, Qwen3.8 or Mistral Small 4 on computers you own or rent. You control where the data lives and pay no usage fees to an AI company. In return you need expensive graphics hardware and the technical skill to run it, and updates are your job. OpenAI says gpt-oss-120b, which has 117 billion parameters (a rough measure of a model's size), fits on a single H100, a data-center graphics card. Running your own makes sense when regulations require that data stay on your premises, or when your volume is so high that owning the hardware costs less than the fees. What it was like for us to run a smaller one is in the next section.
A further-trained model is a base model trained some more on your own examples, and on a narrow job it can beat a much larger general model. Open models make this practical, and Qwen's page points to Unsloth, a free tool for running and training models. Volume decides whether it pays. Airbnb's engineering team tested a small model trained on its own examples against the leading paid models on one real sorting job, and it matched their quality at about a hundredth of the cost. Even so, at October 2026 list prices the paid models came out cheaper below roughly 15 to 30 requests a minute, which is more than 20,000 requests a day. Airbnb's write-up puts it plainly: "For low-volume tasks, or tasks that change week-to-week, a frontier model behind a good prompt is often the right call."
Which models do our own AI agents use?
VantaSoft runs a team of 15 AI agents on one Mac mini in our office, including a chief of staff, a bookkeeper, engineers and the content agent that drafted this guide. Here is what they run on, read from their settings on November 8, 2026.
| Work | Model | Tier |
|---|---|---|
| Everyday work for 14 of the 15 agents | Claude Opus 5.5 | Top tier |
| Editing our proposal drafts | A version of Qwen3.8-27B, on our own Mac mini | Open model |
| Draft images and variations | GPT-Image-2.6 Flare | Fast |
| Images we publish | GPT-Image-2.6 Sunburst | Detailed |
Fourteen of the agents use Claude Opus 5.5, and that choice sits in one shared settings file, so changing the model for the whole team is a one-line edit. Anthropic describes Opus 5.5 as built "for long-running agentic coding and knowledge work," which is close to what our agents spend their days on: writing and checking code, keeping the books, researching and drafting articles like this one. It replaced Claude Opus 5 on September 22, and the list price went down with the switch, from $5 and $25 per million to $4 and $20.
Don't copy this for a first agent. Our agents do long, many-step work all day, which is the work Anthropic points Opus at. A business with one agent sorting email belongs in the tier table above, on a middle or fast model.
One switch taught us to keep the model choice in one place. On October 21, three agents were moved up to Claude Fable 5.1, the top tier at two and a half times Opus 5.5's price, and moved back the same day. Ten of their scheduled jobs had their own model setting, so the move back missed them, and they stayed set to Fable until October 23. On October 26 we removed every job-level setting. All of the team's scheduled jobs now follow the one shared line, and none had a setting of its own when we checked on November 8.
One agent is the exception. Since November 5, the editor that polishes our proposal drafts has run on a version of Qwen3.8-27B, an open model, on the same Mac mini, an M5 Pro with 64 GB of memory. It runs up no usage fees, but it is slower. In our tests on that machine it wrote about 40 units of text a second on a typical reply, roughly 30 words. The first reply of a new conversation waits about a minute while it reads its standing instructions, and later replies skip that wait. We also tried a smaller compressed copy of the model and kept the larger one, which used its tools more reliably.
Before November 5 that editor ran on an online service, after a different provider turned out to accept only 32,000 units of text per request, half of the 64,000 our agent software needs. If you rent an open model from a hosting company, check that limit before you compare names and prices.
Our images follow the tier rule too. Drafts and variations go to OpenAI's GPT-Image-2.6 Flare, which OpenAI lists "for fast iteration," and only the image we publish goes to GPT-Image-2.6 Sunburst, listed "for detailed visual work."
When the model choice doesn't matter
For many business uses, the model is one of the less important decisions you will make. Three other things often matter more.
The first is how you ask. The same model can give very different results depending on its instructions, and careful instructions to a middle-tier model can beat careless ones to a top-tier model. Improve the instructions before you pay for a higher tier.
The second is the quality of your information. An AI model is only as useful as what you give it to read. If your price lists, policies or customer records are outdated, messy or incomplete, no model will give good answers. Cleaning them up pays off whichever model you choose.
The third is how the model fits into your work. Whether it can look things up in your files, use your other software and act on what it finds decides whether it does useful work or stays a novelty, and that shapes results as much as the model does. Our guide to AI agent integration covers how to tell which of your systems are ready to connect.
A simple decision framework
Use these five questions to narrow your choice:
- What is the job? Define exactly what goes in and what should come out. "Sort support emails by urgency" is a job, and "use AI somewhere" is not.
- How accurate does it need to be? If mistakes cost money or trust, start with a top-tier model and decide which results a person checks before they go out (human in the loop for AI agents covers where that check belongs). If mistakes are cheap, as with first drafts or internal summaries, start with a middle-tier model.
- How much will you use it? With the example request above, 10,000 requests a month separates the top and bottom tiers by about $356. A million requests a month separates them by about $35,550. Run your own numbers in the LLM cost calculator.
- Can your data leave your own computers? If not, your options are an open model you run yourself, or an AI company whose data terms and location options meet your compliance requirements. Check the plan you are on, too: Google's pricing page says content sent on the Gemini API's free tier is used to improve its products, and content on the paid tier is not.
- Do you have technical staff who can run AI models? If not, stay with the AI companies' online services, because running and training your own models needs ongoing technical work.
For most businesses starting with AI, the answer is a middle-tier model from one of the big AI companies, time spent on instructions and data quality, and a second look at the model once you have real usage numbers.
Frequently asked questions
Which AI model should I use for my small business?
For most small businesses, a middle-tier model from one of the big AI companies gives the best balance of quality and cost. In November 2026 that means Claude Sonnet 5.5, GPT-6.1 Sol or Gemini 3.8 Flash. Their vendors present them as the balance point of each lineup (Anthropic calls Sonnet 5.5 "the best combination of speed and intelligence"), and you pay only for what you use.
Is it worth paying for a top-tier model?
Top-tier models such as Claude Fable 5.2, Claude Opus 5.5 and GPT-6 Astra are worth the premium when accuracy is critical and mistakes are expensive, as in reviewing legal documents, writing complex software and pulling together research. For routine jobs like sorting email or answering common questions, start with a cheaper tier and move up only if your own tests show you need to.
Which LLM does VantaSoft use for its own AI agents?
Fourteen of our 15 agents run on Claude Opus 5.5, set in one shared file so the whole team changes models with one edit. Our proposal editor runs on a version of Qwen3.8-27B, an open model, on our own Mac mini. For images we use GPT-Image-2.6 Flare for drafts and Sunburst for the final picture.
How often should I re-evaluate my LLM choice?
Review your choice at least once a quarter. Prices for the same tier fall quickly. Anthropic's price list shows Claude Opus 4 and 4.1 at $15 to read and $75 to write per million, Opus 4.5 through Opus 5 at $5 and $25, and Opus 5.5 at $4 and $20. Vendors also retire older models. Google's pricing page says Gemini 2.5 Pro is deprecated and that from December 1, 2026, requests to it go to Gemini 3.8 Flash and are billed at that model's rates. OpenAI keeps a deprecations page listing the ones on their way out.
Can I use multiple LLMs in the same system?
Yes. You can send sorting and tagging to a fast, cheap model, which is the work Anthropic pitches Haiku 5.5 for, and pass harder tasks to a more capable one. Simple work stays cheap, and the expensive model only sees the tasks that need it.
What about open-source models vs. commercial APIs?
Open models (often called open source) such as gpt-oss-120b, Qwen3.8 and Mistral Small 4 can handle real business work, and they are the usual answer when privacy rules keep you from sending data to an AI company. What you pay instead is hardware and upkeep. With specialist engineers on staff, they can cost less than the AI companies' fees at high volume. Without them, the hardware, monitoring and updates often cost more than the fees you avoided.
Start with the use case
Choosing an LLM is a business decision more than a technology decision. The right model meets your accuracy needs, fits your budget at the volume you expect and follows your rules about data.
Start with the job rather than the model, try it with a middle-tier model and adjust your choice once you have real numbers. Expect to change models as vendors release new ones and retire old ones, and keep the choice in one setting so a switch is a quick edit rather than a rebuild. If you are planning your first AI agent, our guide to implementing an AI agent covers the decisions to settle before the build starts.
If you would rather not make these choices yourself, our Agent Service builds an agent around one of your workflows and handles the machine, the model setup and every update after that. Start Automating. It starts with 30 minutes on your workflows.




