There is a new "best model" announcement every few weeks. Charts get shared, rankings shift, and it can feel like you are expected to pick a winner.
For a small or mid-sized business, that is the wrong frame. A ranking tells you which model did best on someone else's questions. You have your own questions: your invoices, your customers, your policies. This post gives you a way to choose between kinds of models, as of November 2026, without leaning on any leaderboard.
It does not name specific models or quote benchmark scores. Those change monthly, and a number out of context will mislead you.
The three kinds of model, simply
- Frontier models are the largest and most capable systems from the big labs. They are usually reached over the internet through a provider's service, and priced by how much text you send and receive.
- Small models have far fewer parameters. They are cheaper and faster to run and can be good at narrow jobs: sorting, extracting fields, short replies.
- Open-weight models are ones whose trained parameters are published so you can run them yourself, on your own hardware or a host you choose. Open-weight is not always the same as open source: the licence may carry conditions, so read it.
These categories overlap. A model can be small and open, or large and closed. The useful question is which properties you need.
Size is not the whole story
Two research results are worth keeping in mind.
In the 2022 InstructGPT paper from OpenAI, human evaluators preferred the outputs of a 1.3 billion parameter model, tuned with human feedback, to those of the 175 billion parameter GPT-3. In the same year, DeepMind's "Training Compute-Optimal Large Language Models" reported that a 70 billion parameter model trained on about four times more data outperformed a 280 billion parameter one trained with the same compute budget, and noted that being smaller makes it cheaper to fine-tune and to run.
Neither result says small always wins. They say how a model is trained and tuned, and what it is trained on, can matter as much as its size. So do not assume bigger is better for your task, and do not assume smaller is.
Six choices to weigh
1. Quality on your task
This is the one that counts most and is hardest to read off a chart. A frontier model is usually better at open-ended reasoning, long documents and tasks with messy instructions. A small model can match it on a narrow, well-defined task, such as tagging an email as quote request, complaint or invoice question.
2. Cost per task
Prices are usually set per amount of text in and out. A model that is cheaper per word but needs three tries is not cheaper. Think in cost per finished task: what does it cost to get one invoice coded, one email drafted, one record matched?
3. Speed
A small model usually responds faster. For a live chat on your site, that can matter. For an overnight reconciliation, it does not.
4. Privacy and where data goes
When you use a hosted model, your text goes to the provider. Check their terms on retention and on whether your inputs are used for training. Running an open-weight model on infrastructure you control keeps data closer to home, at the price of running it yourself. Québec's private-sector privacy law, as amended by Law 25, and similar rules elsewhere can affect where personal information may go; ask your privacy lead or lawyer before sending customer data to any service.
5. Control
With a hosted model, the provider decides when versions change and when old ones are retired. That can change how your workflows behave. An open-weight model you host stays as it is until you change it. Control also means more of the work is yours: updates, security, monitoring.
6. Reliability and tooling
Can the model follow a strict format? Call your tools correctly? Say "I don't know"? These matter more for an agent than a clever turn of phrase.
Measure more than accuracy
The Stanford HELM project ("Holistic Evaluation of Language Models", 2022) made a point that applies to a business: evaluate on more than one measure. Their list was accuracy, calibration, robustness, fairness, bias, toxicity and efficiency, and they argued the trade-offs between them should be visible rather than hidden behind one score.
For your own purposes, a short version:
- Accuracy: is the answer right?
- Consistency: does it give the same answer to the same question tomorrow?
- Failure behaviour: when it is unsure, does it say so or invent something?
- Cost and speed per finished task.
- Safety with your data.
How to test on your own work
1. Collect 30 to 50 real examples from your own business with answers a person already knows: invoices, customer messages, quotes. Include a few awkward ones.
2. Run each candidate model on the same examples, with the same instructions, and score the results against the known answers.
3. Check consistency by running the same example a few times.
4. Note the cost and time per finished task.
5. Re-run when the model or the prompt changes. A set of examples you keep is the cheapest protection you can have against a silent change.
If a small model passes your examples, you probably do not need a bigger one for that task. If it fails on the hard cases only, consider using a small model for the easy majority and a larger one for the rest, with a person reviewing what is uncertain. This is often called routing.
What this means for your business
- Do not pick on a ranking. Pick on your own tasks, with your own test set.
- Match the model to the job. Sorting and extracting rarely need the most capable model. Complex drafting and judgment calls may.
- Put privacy in the first question, not the last. Where does the data go, and who may see it?
- Plan for change. Models are replaced often. Your tests and your records should outlast any one model.
- Keep a person in the loop on what leaves the company. No model choice replaces approval of a message, a price or a payment.
How this shows up at Humonix
We do not make you pick a model. Hue runs on models from outside providers, and AI use is included at the listed volumes in each plan; beyond those, it is billed at the provider's list price with no markup, or you can use your own model key (see Bring your own). Enterprise work, scoped case by case, includes things like model routing and private hosting.
Your facts stay in Know, separate from any model, so changing models does not mean starting over. Anything public, priced or paid waits for a person's approval in Hue. And each feature on a module page is labelled Live, Early access or Coming, so a model's potential is never sold as a feature that runs today. For the architecture, see The Agentic Enterprise.
Sources
- Ouyang et al., "Training language models to follow instructions with human feedback" (arXiv:2203.02155, 2022)
- Hoffmann et al., "Training Compute-Optimal Large Language Models" (arXiv:2203.15556, 2022)
- Liang et al., "Holistic Evaluation of Language Models" (arXiv:2211.09110; TMLR 2023)
- NIST, AI Risk Management Framework and Generative AI Profile (NIST AI 600-1)
This post is general information, not technical or legal advice.